<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0"
     xmlns:atom="http://www.w3.org/2005/Atom"
     xmlns:content="http://purl.org/rss/1.0/modules/content/"
     xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>Claude Code · Loosely Decoupled</title>
    <link>https://looselydecoupled.com/topics/claude-code/</link>
    <description>Claude Code in use: what it leaves on disk, why a plugin loads in one session and not another, and prompts that triage footage.</description>
    <language>en-us</language>
    
    <copyright>© 2026 Emmanuel Tsouris</copyright>
    <lastBuildDate>Tue, 20 Oct 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://looselydecoupled.com/topics/claude-code/index.xml" rel="self" type="application/rss+xml" />
    
    <item>
      <title>What Claude Code Leaves on Your Mac&#39;s Disk, and What Not to Delete</title>
      <link>https://looselydecoupled.com/claude-code-disk-space-worktrees/</link>
      <guid isPermaLink="true">https://looselydecoupled.com/claude-code-disk-space-worktrees/</guid>
      <pubDate>Tue, 20 Oct 2026 00:00:00 +0000</pubDate>
      <dc:creator>Emmanuel Tsouris</dc:creator>
      <category>Disk space</category>
      <category>Claude Code</category>
      
      <description>A read-only prompt makes Claude Code sort its scratch folders, worktrees and session history into in use, recoverable or unknown, with evidence for each.</description>
      <content:encoded><![CDATA[<div class="answer" data-answer="true">
  <p class="answer__label" data-pagefind-ignore>TL;DR</p>
  <p>Do not delete these folders yet. Claude Code&rsquo;s scratch checkouts, its worktrees and its <code>~/.claude</code> directory can hold unpushed commits and session history that exist nowhere else. Paste the prompt below into Claude Code in <a href="https://code.claude.com/docs/en/permission-modes#plan-mode" rel="external noopener" target="_blank">plan mode<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a>. Claude Code measures every large item and every worktree and classifies each as IN USE, RECOVERABLE or UNKNOWN, with a line of evidence. Treat UNKNOWN as the default, and never treat a session transcript as recoverable, because a stopped session can be resumed from it. Read the limits before you act on RECOVERABLE, and approve removals one item at a time; the prompt tells Claude Code to delete nothing until you do.</p>

</div>

<p>On October 20 I measured the folders that Claude Code leaves behind on my Mac again. Size and age won&rsquo;t sort them: a folder that looks abandoned can hold the only copy of a commit, or a session you can still resume. This page is part of a series on <a href="https://looselydecoupled.com/series/developer-mac-disk-cleanup/">cleaning up a developer Mac&rsquo;s disk</a>. If you don&rsquo;t yet know that agents are what fills your disk, <a href="https://looselydecoupled.com/mac-disk-full-developer-caches/">finding what&rsquo;s taking up the space</a> comes first. If the space is a uv cache that agents filled through MCP servers, <a href="https://looselydecoupled.com/uv-cache-huge-mac-uvx-mcp-servers/">that has its own cleanup</a>.</p>
<div class="tested" id="tested">
  <p class="label label--seam">Tested with</p><dl class="tested__rows">
    <div class="tested__row"><dt>macOS</dt> <dd>27.0.1 (26A434), APFS, Claude Code=2.1.304</dd></div>
    <div class="tested__row"><dt>Date</dt> <dd>October 2026</dd></div>
  </dl>
</div>
<h2 id="run-the-prompt">Run the prompt<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#run-the-prompt"
     aria-label="Link to section: Run the prompt" data-pagefind-ignore></a>
</h2>
<p>Paste this into Claude Code in <a href="https://code.claude.com/docs/en/permission-modes#plan-mode" rel="external noopener" target="_blank">plan mode<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a>. Plan mode doesn&rsquo;t limit Bash to read-only commands on its own, and it writes a plan file of its own. The prompt&rsquo;s no-write rule is the guard, so watch the first tool calls; see <a href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#common-failures">Common failures</a>. I ran it in a fresh session started with <code>claude -p &quot;&lt;prompt&gt;&quot; --permission-mode plan</code>. A <code>-p</code> run is one-shot and ends after the report, so I haven&rsquo;t tested the prompt&rsquo;s approval step, which applies only in an interactive session.</p>
<div class="code-block" data-copyable data-copy-label="Text">
  <div class="code-block__head"><span class="code-block__lang">Text</span>
  </div>
  <div class="code-block__scroll">
    <div data-copy-source><div class="highlight"><pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">You are helping me find out how much disk space my coding agents and their git worktrees are using, and which of it is safe to remove. Do not delete, move, or modify anything, and do not write any file, including a plan or notes file; answer in the chat only. Do not run any command that writes. Use the read-only form of git (git --no-optional-locks status) because plain git status can rewrite the index. If I want something removed, you will propose it and wait for my approval.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">1. Measure. Report the size of: (a) the agent temporary scratch area (for tools that use a per-session scratchpad under the system temp folder, list each session&#39;s scratchpad with its size), (b) the agent&#39;s configuration and session-history directory in my home folder, broken down by its largest subdirectories, (c) every git worktree: run `git worktree list` in each repository under my projects folder and give each linked worktree&#39;s path and size. Use du; note that du counts APFS clones at full size, so totals are an upper bound.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">2. Classify every item over 100 MB, and every worktree, as exactly one of:
</span></span><span class="line"><span class="cl">   - IN USE: a running process has its working directory or files there (check with ps and lsof on the process list, not by assumption), or it is the scratchpad of a session that is running now (including this one).
</span></span><span class="line"><span class="cl">   - RECOVERABLE: nothing is running there, `git status` shows no uncommitted changes, no untracked files you cannot regenerate, and every local commit is pushed (compare with the upstream branch); or it is a plain build or dependency folder (node_modules, .build, .venv) that can be rebuilt.
</span></span><span class="line"><span class="cl">   - UNKNOWN: anything else, including when you could not check. A session transcript or history folder is never RECOVERABLE: a stopped session can still be resumed from it, and it cannot be regenerated. A folder holding a worktree whose branch has commits that exist nowhere else is UNKNOWN even if the folder itself is missing.
</span></span><span class="line"><span class="cl">   For each item give the evidence in one line: the process found or not found, the git status result, the unpushed commit count, the last-modified time.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">3. Report as a table: path, size, class, evidence. Then totals per class. State what your checks could not see, for example a session that is stopped but could be resumed, or a worktree with a stash.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">4. Stop. Do not delete anything. Ask which items I want removed, and say for each which tool&#39;s own command you would use (`git worktree remove`, `swift package reset`, deleting the folder).</span></span></code></pre></div></div>
  </div>
</div><p>The prompt avoids Claude Code&rsquo;s names, but I tested it only in Claude Code.</p>
<h2 id="read-the-table-by-class">Read the table by class<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#read-the-table-by-class"
     aria-label="Link to section: Read the table by class" data-pagefind-ignore></a>
</h2>
<p>A session transcript or history folder is never RECOVERABLE: a stopped session can be resumed from it with <code>--resume</code>, and it can&rsquo;t be regenerated. In my run, a running session&rsquo;s transcripts came back IN USE and the rest UNKNOWN. Claude Code already deletes transcripts older than its <code>cleanupPeriodDays</code> setting (see <a href="https://code.claude.com/docs/en/settings-reference#cleanupperioddays" rel="external noopener" target="_blank">the settings reference<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a>); change that setting rather than deleting transcripts by hand.</p>
<p>A worktree whose folder is gone can still matter. A worktree is a second working folder for a repository, and git records it in the repository itself. If the worktree was on a detached HEAD, its HEAD in the repository may be the only reference to its commits, even though the folder is gone. A named branch survives removing the worktree; its unpushed commits are at risk only if you later delete the branch. The prompt classes such a worktree as UNKNOWN.</p>
<p>The report ends with totals per class and a list of what the checks couldn&rsquo;t see. Read the list before the totals.</p>
<h2 id="what-the-report-cannot-see">What the report cannot see<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#what-the-report-cannot-see"
     aria-label="Link to section: What the report cannot see" data-pagefind-ignore></a>
</h2>
<p>Claude Code can&rsquo;t check everything, and the prompt makes it say so. In my run it listed these:</p>
<ul>
<li><strong>A stopped session that could be resumed.</strong> The check finds that no process is running, not whether you&rsquo;ll come back to it.</li>
<li><strong>Short-lived processes.</strong> A process that wasn&rsquo;t running at the moment of the check doesn&rsquo;t show up in <code>ps</code>.</li>
<li><strong>Remote-tracking refs last fetched months ago.</strong> The unpushed count compares local commits with the last-fetched state of the remote, so it can be wrong in either direction. In the test Claude Code didn&rsquo;t run <code>git ls-remote</code>, which contacts the network.</li>
<li><strong>A stash.</strong> Stashes are repository-wide, and a worktree&rsquo;s <code>git status</code> doesn&rsquo;t show them.</li>
<li><strong>Ignored files in clones.</strong> Files that git ignores don&rsquo;t appear in <code>git status</code>, so a RECOVERABLE clone can still hold something you made.</li>
</ul>
<p><code>du</code> has a limit of its own. It counts APFS clones at full size, so every total is an upper bound on what deleting frees. Measure the container&rsquo;s free space before and after, as in <a href="https://looselydecoupled.com/mac-disk-full-developer-caches/">finding what&rsquo;s taking up the space</a>, to see what the Mac actually got back.</p>
<h2 id="approve-removals-one-item-at-a-time">Approve removals one item at a time<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#approve-removals-one-item-at-a-time"
     aria-label="Link to section: Approve removals one item at a time" data-pagefind-ignore></a>
</h2>
<p>The prompt&rsquo;s last step tells Claude Code to stop after the report and ask which items you want removed. For each item it names the tool&rsquo;s own command: <code>git worktree remove</code> for a worktree, <code>swift package reset</code> for a Swift package&rsquo;s build state, or deleting the folder. My test run was one-shot, so I haven&rsquo;t seen this step in an interactive session.</p>
<p>Decide per item, not per class. RECOVERABLE means the checks found nothing to lose, not that the item is certain to be safe, so go through the limits in the previous section first. A line of evidence you can&rsquo;t explain is a reason to refuse that item.</p>
<div class="admonition admonition--warning" role="note">
  <p class="admonition__label">
    <svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#i-triangle-alert"/></svg>
    <span>Warning</span>
  </p>
  <p>Do not let the agent run <code>git worktree prune</code> while a missing worktree in the repository is on a detached HEAD. Pruning clears every missing worktree in a repository, not only the one you have in mind. A detached worktree&rsquo;s HEAD can be the only reference to its commits (<a href="https://git-scm.com/docs/git-worktree" rel="external noopener" target="_blank">git worktree<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a>). Git also removes a missing worktree&rsquo;s records on its own after a while (<code>gc.worktreePruneExpire</code>), so do not leave those commits without a branch. Give each one a branch first: read the commit from <code>git worktree list --porcelain</code> and run <code>git branch keep-&lt;name&gt; &lt;commit&gt;</code>. Remove a worktree whose folder still exists with <code>git worktree remove &lt;path&gt;</code>; removing it without a branch drops the reference exactly as pruning does. The report drew the same line: it called pruning safe for worktrees whose branches existed and unsafe for those on a detached HEAD.</p>
</div><h2 id="what-one-run-found">What one run found<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#what-one-run-found"
     aria-label="Link to section: What one run found" data-pagefind-ignore></a>
</h2>
<p>Numbers from one Mac on October 20. The prompt asks about all my coding agents, so the report also measured a second agent&rsquo;s directory, <code>~/.codex</code>. It was the largest item, and this page doesn&rsquo;t cover cleaning it. Worktrees whose folders were gone still referenced commits.</p>
<div class="table-wrap" data-scrollable data-label="Table: Item">
  <table>
    <thead>
      <tr><th scope="col">Item</th><th scope="col">Size</th><th scope="col">What it was</th></tr>
    </thead>
    <tbody>
      <tr><td>Claude Code temporary scratch area</td><td>2.0 GB</td><td>66 per-session scratchpads; the largest 1.3 GB, the next 246 MB</td></tr>
      <tr><td><code>~/.claude</code></td><td>3.7 GB</td><td>Session transcripts (<code>projects</code>) 2.6 GB, the largest part</td></tr>
      <tr><td>A second agent&rsquo;s directory, <code>~/.codex</code></td><td>7.8 GB</td><td>Packages 2.0 GB, sessions 1.9 GB, a history database 1.5 GB, generated images 1.1 GB</td></tr>
    </tbody>
  </table>
</div>
<p>Claude Code classed the second agent&rsquo;s package folder IN USE, because a running process executed from it. It classed three worktrees on a detached HEAD, whose folders were gone, as UNKNOWN.</p>
<p>The class totals were about 2.6 GB IN USE, 1.5 GB RECOVERABLE and 8.7 GB UNKNOWN; most of the UNKNOWN was the rest of the second agent&rsquo;s directory. The totals cover only what the prompt classifies, items over 100 MB and worktrees, so they don&rsquo;t add up to the table.</p>
<p>I checked the Claude Code sizes against my own <code>du</code>; the <code>~/.codex</code> figures are the agent&rsquo;s. I checked that the second run wrote no plan file. I didn&rsquo;t check each classification line. Treat the classes as the agent&rsquo;s report, not as my measurement.</p>
<p>Claude Code counted 19 linked worktrees across 9 repositories. My own search, to six directory levels, found 16, plus 12 <code>.claude/worktrees</code> directories under my projects folder.</p>
<h2 id="common-failures">Common failures<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#common-failures"
     aria-label="Link to section: Common failures" data-pagefind-ignore></a>
</h2>
<h3 id="plan-file-appears-after-a-read-only-run">Plan file appears after a read-only run<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-disk-space-worktrees/#plan-file-appears-after-a-read-only-run"
     aria-label="Link to section: Plan file appears after a read-only run" data-pagefind-ignore></a>
</h3>
<p>A new file appears in <code>~/.claude/plans</code> after the report. Plan mode writes a plan file by a mechanism of its own, and a prompt that only says not to modify anything doesn&rsquo;t stop it. My first run, with that wording, wrote one. Keep the prompt&rsquo;s line that forbids writing any file, including a plan or notes file; with it, my second run wrote none. After a run, list the folder newest first and check for a file from that run:</p>
<div class="console"><div class="code-block code-block--wrap console__command" data-copyable data-copy-label="command">
    <div class="code-block__head"><span class="code-block__lang">Command</span></div>
    <div class="code-block__scroll"><pre data-copy-source><code><span class="line"><span class="prompt">$ </span><span class="console__token">ls</span> <span class="console__token">-lt</span> <span class="console__token">~/.claude/plans</span></span></code></pre>
    </div>
  </div></div>]]></content:encoded>
    </item>
    
    <item>
      <title>Why Doesn&#39;t My Claude Code Plugin Load in a Cloud Session?</title>
      <link>https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/</link>
      <guid isPermaLink="true">https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/</guid>
      <pubDate>Sat, 17 Oct 2026 00:00:00 +0000</pubDate>
      <dc:creator>Emmanuel Tsouris</dc:creator>
      <category>Claude Code</category>
      <category>DevOps</category>
      
      <description>The install worked in my test cloud session only because it had the plugin&#39;s private repository attached. Sessions without it failed silently.</description>
      <content:encoded><![CDATA[<div class="answer" data-answer="true">
  <p class="answer__label" data-pagefind-ignore>TL;DR</p>
  <p>Claude Code cloud sessions don&rsquo;t install the plugins that a repository&rsquo;s <code>.claude/settings.json</code> turns on. A cloud session&rsquo;s GitHub proxy serves only the repositories attached to that session. A plugin from a private marketplace therefore doesn&rsquo;t load in a cloud session. The plugin&rsquo;s skill can still load. Zip the skill folder. Upload the zip under <strong>Customize</strong> &gt; <strong>Skills</strong> on claude.ai. Then turn the skill on. In a new cloud session with nothing attached, the uploaded skill showed up as <code>anthropic-skills:&lt;skill&gt;</code>. This workaround has two trade-offs. You upload the skill again for every release. The skill also loads in every session on your account. A missing skill doesn&rsquo;t raise an error. Claude still answers, but without the skill&rsquo;s process. To check, ask the session to list its skills.</p>

</div>

<p>I was working on a project spread across several private repositories, and I created a skill I wanted to share between them as a Claude Code plugin. My plan was a private GitHub repository that&rsquo;s both the plugin and its marketplace, so the skill stays in sync on my laptop and in cloud sessions. Most of my work happens in Claude Code on the web, because I like that it keeps the work off my laptop, runs in isolation, and is there on my phone. I tested the install commands by hand in a cloud session, then put them in the environment&rsquo;s setup script, and the test worked. In the next real session the skill never loaded, and I got no error, not even a small one. Here&rsquo;s how I tracked down the cause, what I tried that didn&rsquo;t work, and the two workarounds that did.</p>
<h2 id="skill-doesnt-load-in-a-cloud-session-and-theres-no-error">Skill doesn&rsquo;t load in a cloud session, and there&rsquo;s no error<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#skill-doesnt-load-in-a-cloud-session-and-theres-no-error"
     aria-label="Link to section: Skill doesn’t load in a cloud session, and there’s no error" data-pagefind-ignore></a>
</h2>
<p>The session starts normally. You paste the notes the skill is built for, and Claude gives a reasonable, generic answer instead of following the skill.</p>
<p>That&rsquo;s what makes it easy to miss. When a skill doesn&rsquo;t load, Claude still answers, and nothing hints that a skill should have run. The only clue I had was that the answer was good but didn&rsquo;t follow the skill&rsquo;s steps.</p>
<p>My test had passed only because of how that session was set up. The session I tested in had the skills repository attached as one of its sources. The real sessions had only the repository I was working in. I also never confirmed that the saved setup script still had the install commands in it. Its log only showed output from another installer.</p>
<p>A second session found the cause. I messaged the failing session from the one where I built the plugin and asked it to inspect its own environment: the plugin list, the contents of <code>~/.claude</code>, and the settings files. It reported that no marketplace had ever been added.</p>
<h2 id="cloud-session-cant-reach-a-private-repository-that-isnt-attached">Cloud session can&rsquo;t reach a private repository that isn&rsquo;t attached<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#cloud-session-cant-reach-a-private-repository-that-isnt-attached"
     aria-label="Link to section: Cloud session can’t reach a private repository that isn’t attached" data-pagefind-ignore></a>
</h2>
<p>A cloud session reaches GitHub through a proxy, and the <a href="https://code.claude.com/docs/en/cloud-environments#github-proxy" rel="external noopener" target="_blank">cloud environments documentation<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> says it serves GitHub API requests only for the repositories attached to the session. Git behaved the same way for me. In a fresh session with only the consuming repository attached, I tried to reach the skills repository:</p>
<div class="console"><div class="code-block code-block--wrap console__command" data-copyable data-copy-label="command">
    <div class="code-block__head"><span class="code-block__lang">Command</span></div>
    <div class="code-block__scroll"><pre data-copy-source><code><span class="line"><span class="prompt">$ </span><span class="console__token">git</span> <span class="console__token">ls-remote</span> <span class="console__token">https://github.com/&lt;owner&gt;/&lt;skills-repo&gt;.git</span></span></code></pre>
    </div>
  </div><div class="code-block code-block--wrap console__output">
    <p class="console__output-label">Output</p>
    <div class="code-block__scroll"><pre><code><span class="line"><span class="console__token">fatal:</span> <span class="console__token">could</span> <span class="console__token">not</span> <span class="console__token">read</span> <span class="console__token">Username</span> <span class="console__token">for</span> <span class="console__token">&#39;https://github.com&#39;:</span> <span class="console__token">terminal</span> <span class="console__token">prompts</span> <span class="console__token">disabled</span></span></code></pre>
    </div>
  </div></div><p>No credentials were supplied for it, so git asked for a username and gave up, even though the Claude GitHub App had access to that repository on my account. I didn&rsquo;t run the same command against the attached repository to compare. The same session hit the limit again, this time with nothing to do with plugins: <code>npm ci</code> couldn&rsquo;t fetch the project&rsquo;s private SDK, which lives in another private repository that wasn&rsquo;t attached.</p>
<p>Here&rsquo;s what that means for each way of installing the plugin:</p>
<ul>
<li><strong>Repository settings:</strong> The <a href="https://code.claude.com/docs/en/cloud-environments#what-carries-over-from-your-setup" rel="external noopener" target="_blank">documentation<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> lists plugins and marketplaces declared in a repo&rsquo;s <code>.claude/settings.json</code> as not available in cloud sessions. I never put the plugin there, so this is what the docs say, not something I tested.</li>
<li><strong>Setup-script install:</strong> I never confirmed my install commands ran, and the log was empty. If the script can only reach attached repositories, like the session, it had nothing to fetch. I haven&rsquo;t tested that.</li>
<li><strong>Plugin installed mid-session:</strong> In one session, installing from an attached copy left the plugin enabled but not loaded, and the cloud UI has <code>/reload-skills</code> but no <code>/reload-plugins</code>. In an earlier session the same install did make the skill appear. I don&rsquo;t know what was different.</li>
</ul>
<h2 id="plugin-added-on-claudeai-never-reaches-the-session">Plugin added on claude.ai never reaches the session<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#plugin-added-on-claudeai-never-reaches-the-session"
     aria-label="Link to section: Plugin added on claude.ai never reaches the session" data-pagefind-ignore></a>
</h2>
<p>Your claude.ai account has a plugins page too. Under <strong>Customize</strong> &gt; <strong>Plugins</strong> &gt; <strong>Add</strong> &gt; <strong>Add marketplace</strong> &gt; <strong>Add from a repository</strong>, claude.ai takes a private GitHub repository once the Claude GitHub App can read it. The <a href="https://claude.com/docs/plugins/overview" rel="external noopener" target="_blank">plugins overview<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> says a plugin you add there downloads to Claude Code as a synced plugin, which the <a href="https://code.claude.com/docs/en/plugins/loading#synced-plugins" rel="external noopener" target="_blank">plugin loading reference<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> names <code>&lt;name&gt;@synced</code>. That reference says synced plugins load in Cowork and in signed-in terminal sessions. It doesn&rsquo;t mention cloud sessions, although mine did get Anthropic&rsquo;s synced plugins.</p>
<figure><img src="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/add-marketplace_hu_cc27f6804a5ca6de.png" alt="The Add marketplace dialog on claude.ai, with a URL field that takes a GitHub owner/repo or a Git repository URL." width="704" height="264" loading="lazy" decoding="async"><figcaption>Customize, Plugins, Add, Add marketplace, Add from a repository.</figcaption>
</figure>
<p>The marketplace synced and the plugin showed as added, with no error. In the ten minutes I watched, it never showed up anywhere:</p>
<ul>
<li><strong>Cloud session, started five minutes after the add:</strong> The synced manifest under <code>~/.claude/plugins/synced/</code> listed 10 plugins, all from Anthropic&rsquo;s <code>knowledge-work-plugins</code> marketplace. Mine wasn&rsquo;t there.</li>
<li><strong>Terminal on my Mac, started ten minutes after the add:</strong> It printed <code>Plugins changed. Run /reload-plugins to activate.</code>, and the rewritten manifest listed 4 plugins from the same Anthropic marketplace. Mine wasn&rsquo;t there either.</li>
<li><strong><code>claude plugin marketplace list</code>:</strong> Its <code>From claude.ai:</code> section listed only <strong>My Uploads</strong> and the Anthropic Directory.</li>
</ul>
<p>One caveat: my Mac already had the same plugin installed from the command line, which may have affected that result. The cloud session had no local copy, so its result is clean. <strong>Add</strong> &gt; <strong>Upload plugin</strong> is a second way to put a plugin on your account. I didn&rsquo;t try it.</p>
<p>An open issue, <a href="https://github.com/anthropics/claude-code/issues/92031" rel="external noopener" target="_blank">anthropics/claude-code#92031<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a>, describes the same thing from another account: plugins installed on claude.ai never reach cloud sessions, while skills enabled on the account do. That&rsquo;s the route in the next section.</p>
<h2 id="upload-the-skill-to-your-claudeai-account">Upload the skill to your claude.ai account<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#upload-the-skill-to-your-claudeai-account"
     aria-label="Link to section: Upload the skill to your claude.ai account" data-pagefind-ignore></a>
</h2>
<p>A skill doesn&rsquo;t need to be a plugin to reach a cloud session. The <a href="https://code.claude.com/docs/en/skills#skills-in-cowork-and-cloud-sessions" rel="external noopener" target="_blank">skills documentation<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> says cloud sessions load the skills enabled for your claude.ai account, synced at session start.</p>
<ol>
<li>
<p>On your machine, zip the skill folder, with its <code>SKILL.md</code> inside the folder. To build the zip from a release tag in the skills repository, run:</p>
<div class="console"><div class="code-block code-block--wrap console__command" data-copyable data-copy-label="command">
    <div class="code-block__head"><span class="code-block__lang">Command</span></div>
    <div class="code-block__scroll"><pre data-copy-source><code><span class="line"><span class="prompt">$ </span><span class="console__token">git</span> <span class="console__token">archive</span> <span class="console__token">--format=zip</span> <span class="console__token">--prefix=&lt;skill&gt;/</span> <span class="console__token">-o</span> <span class="console__token">&lt;skill&gt;.zip</span> <span class="console__token">\</span>
</span><span class="line">    <span class="console__token">&lt;tag&gt;:plugins/&lt;plugin&gt;/skills/&lt;skill&gt;</span></span></code></pre>
    </div>
  </div></div></li>
<li>
<p>On claude.ai, go to <strong>Customize</strong> &gt; <strong>Skills</strong> &gt; <strong>Add</strong> &gt; <strong>Upload skill</strong>, upload the zip, and turn the skill on.</p>
</li>
<li>
<p>Start a new cloud session, and ask Claude to list its skills.</p>
</li>
</ol>
<figure><img src="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/upload-skill_hu_769c3e4dfaeb44d3.png" alt="The Upload a skill page on claude.ai. It takes a .md, .zip or .skill file, and a security scan runs on upload." width="1030" height="480" loading="lazy" decoding="async"><figcaption>Customize, Skills, Add, Upload skill.</figcaption>
</figure>
<p>On upload, claude.ai previews the skill&rsquo;s name and description and runs a security scan. In a new cloud session with only the consuming repository attached, the skill showed up as <code>anthropic-skills:&lt;skill&gt;</code>. I checked that in one session. My test of Claude actually using the skill, in the section on confirming it loaded, ran on the copied version from the next section, which has the same files.</p>
<p>There are two costs. claude.ai numbers the skill&rsquo;s versions v1, v2 and so on, not by your git tags, and each release means another upload. It&rsquo;s an account skill, so it loads in every session on the account, not only the repositories that use it. That includes terminal sessions: on my Mac it appeared beside the same skill from the installed plugin, so the session offered it twice. Uninstall the local plugin, or turn the account skill off, to keep one.</p>
<h2 id="copy-the-skill-into-the-cloud-session">Copy the skill into the cloud session<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#copy-the-skill-into-the-cloud-session"
     aria-label="Link to section: Copy the skill into the cloud session" data-pagefind-ignore></a>
</h2>
<p>If a session needs a specific release, copy the skill in from the skills repository instead. The <a href="https://code.claude.com/docs/en/skills" rel="external noopener" target="_blank">skills documentation<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> says Claude Code watches <code>~/.claude/skills/</code> and picks up a skill added there during the session, without a restart.</p>
<ol>
<li>
<p>Start the cloud session with two repositories attached: the one you&rsquo;re working in and the skills repository.</p>
</li>
<li>
<p>Ask Claude where the skills repository was cloned. It was <code>/home/user/&lt;skills-repo&gt;</code> in my sessions.</p>
</li>
<li>
<p>Ask Claude to copy the skill folder into your user skills directory, changing nothing else:</p>
<div class="console"><div class="code-block code-block--wrap console__command" data-copyable data-copy-label="command">
    <div class="code-block__head"><span class="code-block__lang">Command</span></div>
    <div class="code-block__scroll"><pre data-copy-source><code><span class="line"><span class="prompt">$ </span><span class="console__token">mkdir</span> <span class="console__token">-p</span> <span class="console__token">~/.claude/skills</span>
</span><span class="line"><span class="prompt">$ </span><span class="console__token">cp</span> <span class="console__token">-R</span> <span class="console__token">/home/user/&lt;skills-repo&gt;/plugins/&lt;plugin&gt;/skills/&lt;skill&gt;</span> <span class="console__token">\</span>
</span><span class="line">    <span class="console__token">~/.claude/skills/&lt;skill&gt;</span></span></code></pre>
    </div>
  </div></div></li>
<li>
<p>Ask Claude to list its skills. If <code>~/.claude/skills/</code> didn&rsquo;t exist when the session started, run <code>/reload-skills</code> first, because Claude Code wasn&rsquo;t watching it.</p>
</li>
</ol>
<figure><img src="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/reload-skills_hu_873f416ef627492b.png" alt="The prompt box of a cloud session with /reload-skills typed and its description: pick up skills added or changed on disk during this session." width="845" height="140" loading="lazy" decoding="async">
</figure>
<p>The skill loads as plain <code>&lt;skill&gt;</code>, with no prefix. To copy a specific release, run <code>git -C /home/user/&lt;skills-repo&gt; fetch --tags</code> and <code>git -C /home/user/&lt;skills-repo&gt; checkout &lt;tag&gt;</code> before the copy. I haven&rsquo;t tested that step. The copy doesn&rsquo;t carry over to other sessions.</p>
<p>If every session in a repository needs the skill, commit a copy under that repository&rsquo;s <code>.claude/skills/</code> instead. The documentation lists the repository&rsquo;s own skills as available in cloud sessions. The cost is a copy in every repository, which is what the plugin was meant to avoid. I built it once, checked it against the release tag, and didn&rsquo;t merge it.</p>
<p>Both routes bring over the plugin&rsquo;s skills only. Hooks, agents and MCP servers in the plugin don&rsquo;t load this way, and I haven&rsquo;t found a way to load them in a cloud session without <code>/reload-plugins</code>.</p>
<h2 id="confirm-the-skill-loaded">Confirm the skill loaded<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#confirm-the-skill-loaded"
     aria-label="Link to section: Confirm the skill loaded" data-pagefind-ignore></a>
</h2>
<p>There are two checks. The session&rsquo;s skill list shows the skill loaded. Asking a question the skill is built for, without naming the skill, shows that Claude actually uses it. My first test was a request to look over a project and say where it could be better. The session announced the skill&rsquo;s first step and worked through its process. If the answer is generic, ask which skill it used.</p>
<p>For the skill list, ask Claude to list its skills, and look for <code>&lt;skill&gt;</code> or <code>anthropic-skills:&lt;skill&gt;</code>.</p>
<p>A plugin that shows as enabled in <code>claude plugin list</code> isn&rsquo;t proof. In one of my sessions it was enabled and never loaded. A folder in <code>~/.claude/skills/</code> isn&rsquo;t proof either, until the session lists it.</p>
<h2 id="do-this-with-a-coding-agent">Do this with a coding agent<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#do-this-with-a-coding-agent"
     aria-label="Link to section: Do this with a coding agent" data-pagefind-ignore></a>
</h2>
<p>What found my cause was an agent inspecting its own environment. Paste this into the coding agent in the cloud session you want to check. Run it once in a session where the skill works and once where it doesn&rsquo;t, then compare the two reports. Replace the placeholders first.</p>
<div class="code-block code-block--wrap" data-copyable data-copy-label="Text">
  <div class="code-block__head"><span class="code-block__lang">Text</span>
  </div>
  <div class="code-block__scroll">
    <div data-copy-source><div class="highlight"><pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">Inspect your own environment, read-only, and report whether a skill is available and where it came from. Change nothing: do not install, enable, disable, add, remove, copy, or edit anything, and do not print any token, key, or password you come across.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">Skill name: &lt;skill-name&gt;
</span></span><span class="line"><span class="cl">Plugin name: &lt;plugin-name&gt;
</span></span><span class="line"><span class="cl">Skills repository: &lt;owner&gt;/&lt;skills-repo&gt;
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">1. Print the Claude Code version.
</span></span><span class="line"><span class="cl">2. List the installed plugins with their version, scope, and status. Note which ones are &lt;name&gt;@synced.
</span></span><span class="line"><span class="cl">3. List the plugin marketplaces that have been added.
</span></span><span class="line"><span class="cl">4. List ~/.claude/skills/ and ~/.claude/plugins/, including plugins/synced/.
</span></span><span class="line"><span class="cl">5. List the repositories attached to this session, and say whether the skills repository is one of them.
</span></span><span class="line"><span class="cl">6. Run git ls-remote https://github.com/&lt;owner&gt;/&lt;skills-repo&gt;.git and quote any error verbatim.
</span></span><span class="line"><span class="cl">7. Report whether this repository&#39;s .claude/settings.json has enabledPlugins or extraKnownMarketplaces entries.
</span></span><span class="line"><span class="cl">8. Report whether a setup script log exists for this session and whether it contains output from a plugin install.
</span></span><span class="line"><span class="cl">9. List the skills available to this session, and say whether &lt;skill-name&gt; is among them and under what name.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">Then answer in this form:
</span></span><span class="line"><span class="cl">- Skill &lt;skill-name&gt; available: yes or no, and under what name
</span></span><span class="line"><span class="cl">- Plugin &lt;plugin-name&gt; installed: yes or no, and whether it is loaded or only enabled
</span></span><span class="line"><span class="cl">- Synced plugins: the names, or none
</span></span><span class="line"><span class="cl">- Skills repository attached: yes or no
</span></span><span class="line"><span class="cl">- Skills repository reachable with git: yes, or the error
</span></span><span class="line"><span class="cl">- Plugin settings in the repository: the entries, or none
</span></span><span class="line"><span class="cl">- Setup script install output: found, not found, or no log
</span></span><span class="line"><span class="cl">- Anything you could not check, and why</span></span></code></pre></div></div>
  </div>
</div><h2 id="common-failures">Common failures<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#common-failures"
     aria-label="Link to section: Common failures" data-pagefind-ignore></a>
</h2>
<h3 id="git-fails-with-terminal-prompts-disabled">Git fails with <code>terminal prompts disabled</code><a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#git-fails-with-terminal-prompts-disabled"
     aria-label="Link to section: Git fails with terminal prompts disabled" data-pagefind-ignore></a>
</h3>
<div class="console console--error"><div class="code-block code-block--wrap console__output">
    <p class="console__output-label">Error</p>
    <div class="code-block__scroll"><pre><code><span class="line"><span class="console__token">fatal:</span> <span class="console__token">could</span> <span class="console__token">not</span> <span class="console__token">read</span> <span class="console__token">Username</span> <span class="console__token">for</span> <span class="console__token">&#39;https://github.com&#39;:</span> <span class="console__token">terminal</span> <span class="console__token">prompts</span> <span class="console__token">disabled</span></span></code></pre>
    </div>
  </div></div><p>In a Claude Code cloud session, git prints this for a private repository that isn&rsquo;t attached to the session, because the GitHub proxy supplies no credentials for it. A mistyped repository name probably produces the same prompt. Check the name, then attach the repository or start a session with it attached. The documentation describes the proxy&rsquo;s answer to a GitHub API request for an unattached repository as a 403 whose message starts <code>GitHub access to</code> and contains <code>is not enabled for this session</code>.</p>
<h3 id="setup-script-install-leaves-no-trace">Setup-script install leaves no trace<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#setup-script-install-leaves-no-trace"
     aria-label="Link to section: Setup-script install leaves no trace" data-pagefind-ignore></a>
</h3>
<p>First check that the saved setup script still contains the install commands. Mine was never checked. If it does, a private skills repository that isn&rsquo;t attached is a likely cause. The <a href="https://code.claude.com/docs/en/cloud-environments#environment-caching" rel="external noopener" target="_blank">documentation<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> also says later sessions can start from a cached snapshot and skip the setup script, so a missing log line doesn&rsquo;t prove the script failed. Check the session itself with <code>claude plugin list</code>.</p>
<h3 id="enabledplugins-in-claudesettingsjson-does-nothing-in-a-cloud-session"><code>enabledPlugins</code> in <code>.claude/settings.json</code> does nothing in a cloud session<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#enabledplugins-in-claudesettingsjson-does-nothing-in-a-cloud-session"
     aria-label="Link to section: enabledPlugins in .claude/settings.json does nothing in a cloud session" data-pagefind-ignore></a>
</h3>
<p>A cloud session doesn&rsquo;t install the plugins a repository turns on under <code>enabledPlugins</code>. The <a href="https://code.claude.com/docs/en/plugins/loading#plugins-shared-through-a-repository" rel="external noopener" target="_blank">plugin loading reference<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> says it doesn&rsquo;t add the marketplaces under <code>extraKnownMarketplaces</code> either, because that needs the workspace trust dialog, which a cloud session never shows. Those keys still work for local sessions. The open issue <a href="https://github.com/anthropics/claude-code/issues/78119" rel="external noopener" target="_blank">anthropics/claude-code#78119<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a> tracks this, and a comment there reports that plugins declared in organization-managed settings don&rsquo;t install in cloud sessions either. I can&rsquo;t test that on an individual account.</p>
<h3 id="plugin-is-enabled-but-the-skill-isnt-listed">Plugin is enabled but the skill isn&rsquo;t listed<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#plugin-is-enabled-but-the-skill-isnt-listed"
     aria-label="Link to section: Plugin is enabled but the skill isn’t listed" data-pagefind-ignore></a>
</h3>
<p>A plugin installed after the session starts loads only when Claude Code restarts or runs <code>/reload-plugins</code>, and the cloud UI has no <code>/reload-plugins</code>. Upload the skill to your claude.ai account, or copy it into <code>~/.claude/skills/</code>.</p>
<h3 id="install-worked-in-a-test-session-and-fails-in-the-real-one">Install worked in a test session and fails in the real one<a class="heading-anchor" href="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/#install-worked-in-a-test-session-and-fails-in-the-real-one"
     aria-label="Link to section: Install worked in a test session and fails in the real one" data-pagefind-ignore></a>
</h3>
<p>What&rsquo;s attached to a session decides what it can reach. Test in a session set up like the one that will use the skill, with the same repositories attached and nothing more.</p>
<p><strong>Still open:</strong> a way to load a whole private plugin, with its hooks, agents and MCP servers, into a cloud session without attaching its repository. Its skills have a route. I haven&rsquo;t found one for the rest. If you&rsquo;ve found a way that works, I&rsquo;d like to hear about it. Reply to any newsletter issue.</p>
<figure><img src="https://looselydecoupled.com/claude-code-plugin-not-loading-cloud-session/cover_hu_22d45f9a47861538.png" alt="A four-panel comic. Panel one: an engineer at a laptop says “Run the skill.” to a robot. Panel two: the robot instantly holds out a neat stack of paper. It says “Done!” Panel three: the engineer, reading the pages with a worried frown, says “That’s not the skill.” while the robot beams. Panel four: the robot, now wearing a small fake moustache, hands on hips, says “No. It’s my impression of it.”" width="1276" height="530" loading="lazy" decoding="async"><figcaption>I got no error, not even a small one.</figcaption>
</figure>
<div class="tested" id="tested">
  <p class="label label--seam">Tested with</p><dl class="tested__rows">
    <div class="tested__row"><dt>Claude Code</dt> <dd>2.1.302</dd></div>
    <div class="tested__row"><dt>Account</dt> <dd>individual claude.ai plan</dd></div>
    <div class="tested__row"><dt>Cloud session</dt> <dd>claude.ai/code</dd></div>
    <div class="tested__row"><dt>Terminal</dt> <dd>macOS, signed in with a claude.ai account</dd></div>
    <div class="tested__row"><dt>Date</dt> <dd>October 2026</dd></div>
  </dl>
</div>
]]></content:encoded>
    </item>
    
    <item>
      <title>How to Sort Months of Dashcam Footage with Claude Code</title>
      <link>https://looselydecoupled.com/triage-dashcam-footage-claude-code/</link>
      <guid isPermaLink="true">https://looselydecoupled.com/triage-dashcam-footage-claude-code/</guid>
      <pubDate>Sun, 27 Sep 2026 00:00:00 +0000</pubDate>
      <dc:creator>Emmanuel Tsouris</dc:creator>
      <category>Dashcam</category>
      <category>Claude Code</category>
      
      <description>Five copy-paste Claude Code prompts sort months of Rivian R1T dashcam footage by GPS and keyframes, each with the check that caught a real failure.</description>
      <content:encoded><![CDATA[<div class="answer" data-answer="true">
  <p class="answer__label" data-pagefind-ignore>TL;DR</p>
  <p>Give Claude Code one prompt per stage: survey, move, find, cut, publish. Each
prompt states what the agent can&rsquo;t touch. Each prompt also makes three demands.
The first demand is to run on one drive first. The second is to break its own check on purpose and confirm that the check fails. The third is to list what it couldn&rsquo;t verify. The agent
reads the GPS tag and the keyframes instead of the pixels. The agent then proposes
candidates. You confirm every pick on a contact sheet before anything is cut,
moved, or deleted.</p>

</div>

<p>Driving from July to September filled the truck&rsquo;s 2 TB dashcam drive, a
<a href="https://www.amazon.com/dp/B08HN37XC1?tag=loosely-decoupled-20" rel="external noopener sponsored" target="_blank">SanDisk Extreme Portable SSD<svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#i-external-link"/></svg><span class="visually-hidden"> (opens in a new tab)</span></a>
(affiliate link): 240 drive folders,
1,958 GB, five cameras recording on every trip. I wanted to keep only what might
make a video, and the backup drive couldn&rsquo;t hold all of it. Watching it wasn&rsquo;t an
option. I directed Claude Code in plain language instead. It sorted the drives
by where they went using GPS, found the I-90 floating bridge by the color of the tunnel
lights, cut synced clips, and rendered a hyperlapse of every drive.</p>
<figure><img src="https://looselydecoupled.com/triage-dashcam-footage-claude-code/five-cameras-floating-bridge_hu_c7373bb90be2c6e0.jpg" alt="Five fisheye views from the same second on the I-90 floating bridge: front, rear, both sides and the Gear Guard camera over the truck bed. A red car passing behind appears in three views at once." width="1600" height="853" loading="lazy" decoding="async"><figcaption>The five cameras at one second of a crossing.</figcaption>
</figure>
<p>Almost every failure in that work was caught by a check: a size comparison, a
frame count, a positive control, a contact sheet, or me. One failure was caught by
nothing for hours. The prompts in this guide are general versions of the ones I
typed, rewritten for your own footage, and each one carries the guardrail that
the failure taught.</p>
<h2 id="signals-already-in-the-files">Signals already in the files<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#signals-already-in-the-files"
     aria-label="Link to section: Signals already in the files" data-pagefind-ignore></a>
</h2>
<p>A Rivian writes one folder per drive and one MP4 per camera per 15-minute segment.
Two things in those files do most of the work, and neither needs anyone to watch
the video:</p>
<ul>
<li><strong>GPS.</strong> The container&rsquo;s <code>location</code> tag holds one fix per segment, taken at its
start. That&rsquo;s the only GPS in the file; there&rsquo;s no per-second track.</li>
<li><strong>Keyframes.</strong> Every camera writes one keyframe per second. Decoding only the
keyframes, downscaled to 64 × 48, gives a per-second record of motion and color
without decoding the frames between them.</li>
</ul>
<p>Read the tag from any clip:</p>
<div class="console"><div class="code-block code-block--wrap console__command" data-copyable data-copy-label="command">
    <div class="code-block__head"><span class="code-block__lang">Command</span></div>
    <div class="code-block__scroll"><pre data-copy-source><code><span class="line"><span class="prompt">$ </span><span class="console__token">ffprobe</span> <span class="console__token">-v</span> <span class="console__token">error</span> <span class="console__token">-show_entries</span> <span class="console__token">format_tags=location</span> <span class="console__token">-of</span> <span class="console__token">default=nw=1:nk=1</span> <span class="console__token">&lt;clip&gt;.mp4</span></span></code></pre>
    </div>
  </div></div><p>It prints a signed latitude immediately followed by a signed longitude, such as
<code>47.589462-122.2655</code>. Some tags end in <code>/</code> and most don&rsquo;t, and <code>0,0</code> means the truck
had no fix. <a href="https://looselydecoupled.com/rivian-dashcam-gps-location-ffprobe/">Read the GPS location from Rivian dashcam video with
ffprobe</a> covers every spelling. Other
dashcams store GPS differently, which is why the first prompt asks the agent to
find out rather than assume.</p>
<h2 id="tools-and-rules-for-every-stage">Tools and rules for every stage<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#tools-and-rules-for-every-stage"
     aria-label="Link to section: Tools and rules for every stage" data-pagefind-ignore></a>
</h2>
<p>You need:</p>
<ul>
<li>Claude Code, run in a terminal on the machine that has the footage mounted</li>
<li>The <code>ffmpeg</code> and <code>ffprobe</code> commands on the <code>PATH</code></li>
<li>The <code>exiftool</code> command, for the privacy stage</li>
<li>A second drive with room for whatever you keep</li>
</ul>
<p>Every prompt carries three rules, worded for its stage. They&rsquo;re repeated on
purpose, so that a prompt pasted on its own still carries them:</p>
<ul>
<li><strong>One drive first.</strong> A bug found on the first drive costs one drive&rsquo;s time. The
same bug found late in a batch costs every drive before it.</li>
<li><strong>Break the check on purpose.</strong> A check that has never failed proves nothing.
The agent damages its own input or output, confirms the check catches it, and
then restores it.</li>
<li><strong>Say what couldn&rsquo;t be verified.</strong> Matching file sizes aren&rsquo;t matching content,
and a report that doesn&rsquo;t say so reads as if they were.</li>
</ul>
<!-- style-check: procedure -->
<h2 id="survey-every-drive-without-changing-anything">Survey every drive without changing anything<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#survey-every-drive-without-changing-anything"
     aria-label="Link to section: Survey every drive without changing anything" data-pagefind-ignore></a>
</h2>
<p>In Claude Code, from the folder that holds the drive folders, paste:</p>
<!-- style-check: off -->
<div class="code-block code-block--wrap" data-copyable data-copy-label="Text">
  <div class="code-block__head"><span class="code-block__lang">Text</span>
  </div>
  <div class="code-block__scroll">
    <div data-copy-source><div class="highlight"><pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">The folder &lt;footage-folder&gt; holds video from my truck&#39;s dashcam: one subfolder per drive, and one MP4 per camera per segment. Survey it without changing anything.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">For every drive folder, use ffprobe to read each segment&#39;s duration and GPS location tag. Report each drive&#39;s start time, length, GPS points, the distance between its two farthest points, and its size on disk. Ignore any ._ files macOS has left beside the clips on a FAT32 or exFAT drive.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">Rules:
</span></span><span class="line"><span class="cl">- Read only. Do not move, rename, re-encode, or delete anything.
</span></span><span class="line"><span class="cl">- Run on one drive first and show me the result before you run the rest.
</span></span><span class="line"><span class="cl">- The location tag is written in more than one format. Parse it permissively, and treat 0,0 as no fix, not as a place.
</span></span><span class="line"><span class="cl">- Before you run the whole batch, feed your parser a malformed tag on purpose and confirm that it reports the drive instead of skipping it silently.
</span></span><span class="line"><span class="cl">- Cache everything you read in one file, so that later steps don&#39;t read the drives again.
</span></span><span class="line"><span class="cl">- Then propose a way to sort the drives into routine and unusual, and wait. I&#39;ll tell you my regular routes.
</span></span><span class="line"><span class="cl">- Finish with a list of the drives you couldn&#39;t read and anything you couldn&#39;t verify.</span></span></code></pre></div></div>
  </div>
</div><!-- style-check: on -->
<p><strong>What went wrong here.</strong> I asked for road trips, and the agent ranked drives by
distance. The ranking itself showed there were none: no drive in the batch went
far. Distance was the wrong question, and only I could supply the right one: the
corridor across the I-90 bridges that most of the drives used. With that, the
classification became a corridor test. A drive whose every GPS point falls
inside a box around that corridor is a corridor drive; anything else is worth a
look. The test found 168 corridor drives, 1,594 GB, and 14 off the corridor
worth keeping,
171 GB. That&rsquo;s why the prompt ends by
waiting for you. Your routes are context the agent can&rsquo;t get from the files.</p>
<p>The parser also failed on its first real run, on a tag that ended in <code>/</code>. The
fix was a pattern that finds two signed numbers wherever they sit.
Expect the same from your own camera&rsquo;s metadata, and test both spellings.</p>
<h2 id="move-the-drives-you-keep">Move the drives you keep<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#move-the-drives-you-keep"
     aria-label="Link to section: Move the drives you keep" data-pagefind-ignore></a>
</h2>
<p>Decide which drives to keep from the survey, then paste:</p>
<!-- style-check: off -->
<div class="code-block code-block--wrap" data-copyable data-copy-label="Text">
  <div class="code-block__head"><span class="code-block__lang">Text</span>
  </div>
  <div class="code-block__scroll">
    <div data-copy-source><div class="highlight"><pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">Move these drives from &lt;footage-folder&gt; to &lt;backup-drive&gt;/&lt;kept-folder&gt;: &lt;list of drive folders&gt;.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">Rules:
</span></span><span class="line"><span class="cl">- Before you copy anything, compare the total size with the free space on &lt;backup-drive&gt;, and stop if it won&#39;t fit.
</span></span><span class="line"><span class="cl">- Copy first. After each drive, compare every file&#39;s size on the destination with the source.
</span></span><span class="line"><span class="cl">- Do not delete anything from the source until I say so. Show me the drives whose sizes all match, and wait.
</span></span><span class="line"><span class="cl">- When I confirm, remove the source files with a tool that removes only what it has copied, such as rsync --remove-source-files. Do not rely on shell-specific glob features.
</span></span><span class="line"><span class="cl">- Skip any folder that another process, such as Finder, is still writing to or moving.
</span></span><span class="line"><span class="cl">- Report the empty folders that a move leaves behind. Do not delete them, and make every later step skip them.
</span></span><span class="line"><span class="cl">- Move one drive first and show me the result.
</span></span><span class="line"><span class="cl">- Before the batch, truncate a copy in a scratch folder and confirm your size check reports it.
</span></span><span class="line"><span class="cl">- Say plainly that a size match isn&#39;t a content match, and list anything you couldn&#39;t verify.</span></span></code></pre></div></div>
  </div>
</div><!-- style-check: on -->
<div class="admonition admonition--warning" role="note">
  <p class="admonition__label">
    <svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#i-triangle-alert"/></svg>
    <span>Warning</span>
  </p>
  <p>Do not let an agent delete source footage in the same step that copies it. Copy,
verify on the destination, and approve the removal yourself. A move that copies
and then fails to delete costs disk space; a move that deletes and then fails to
copy costs the footage.</p>
</div><p><strong>What went wrong here.</strong> A queued move used a <code>zsh</code> glob qualifier that the
shell Claude Code runs commands in doesn&rsquo;t enable. It would have copied every file
and deleted none. Inspecting the running process caught it, and the move switched
to <code>rsync --remove-source-files</code>, which removes only what it copied. Later, a
cleanup command of the form <code>rm -f $B/*.jpg</code> was blocked by Claude Code&rsquo;s own
guard, because an unset <code>$B</code> would expand the path to <code>/*</code>. The cleanup wasn&rsquo;t
needed, and the guard was right to stop it.</p>
<p>The moves also left empty drive folders behind. Nothing failed because of them
until the hyperlapse run, in a later stage, which is why this prompt asks for
them to be reported now.</p>
<h2 id="find-a-landmark-from-computed-signals">Find a landmark from computed signals<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#find-a-landmark-from-computed-signals"
     aria-label="Link to section: Find a landmark from computed signals" data-pagefind-ignore></a>
</h2>
<p>GPS places a drive to within a 15-minute segment, which is too coarse to find a
bridge crossing. The picture has to do it, and the cheap way to read the picture
is to compute a few numbers per second once and reuse them. Paste:</p>
<!-- style-check: off -->
<div class="code-block code-block--wrap" data-copyable data-copy-label="Text">
  <div class="code-block__head"><span class="code-block__lang">Text</span>
  </div>
  <div class="code-block__scroll">
    <div data-copy-source><div class="highlight"><pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">I want the moment in each drive where the truck passes &lt;landmark&gt;. The GPS has one point per segment, which is too coarse, so find it from the picture.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">1. From keyframes only, downscaled to 64x48, compute per-second signals for each camera: the mean change in green from the previous keyframe, and the mean red, green, and blue. Cache them per drive and per camera, so that no later step decodes video again.
</span></span><span class="line"><span class="cl">2. Propose a signal that marks &lt;landmark&gt;, and explain why it should separate the landmark from things that look like it.
</span></span><span class="line"><span class="cl">3. Propose candidates. Do not pick for me.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">Rules:
</span></span><span class="line"><span class="cl">- Read only on &lt;footage-folder&gt;. Write only under &lt;work-folder&gt;.
</span></span><span class="line"><span class="cl">- When you extract keyframes as raw frames, make sure ffmpeg doesn&#39;t duplicate them back up to the source frame rate.
</span></span><span class="line"><span class="cl">- Before you scan everything, calibrate the signal on a drive where I know the landmark is present and a drive where I know it isn&#39;t. If both read the same, suspect your extraction before the signal.
</span></span><span class="line"><span class="cl">- Run on one drive first.
</span></span><span class="line"><span class="cl">- Build a contact sheet with one row per drive, labeled with the drive&#39;s name, and one frame per candidate in time order. Write a picks file with your best guess for each drive, which I can edit. Then wait for me.
</span></span><span class="line"><span class="cl">- Ignore the first and last few minutes of each drive, where the truck is parking or leaving.
</span></span><span class="line"><span class="cl">- If a result contradicts something already confirmed, check your tool before you report the contradiction: extract the same frame two ways and compare them.
</span></span><span class="line"><span class="cl">- List the drives with no candidate, and anything you couldn&#39;t verify.</span></span></code></pre></div></div>
  </div>
</div><!-- style-check: on -->
<p>For a tunnel, the signal turned out to be color. Auto-exposure keeps a tunnel&rsquo;s
brightness close to daylight&rsquo;s, but sodium lighting tips the front camera&rsquo;s red/blue
ratio above 1.0. A run of 25–150 s above that ratio, away from the first and last
3 minutes of the drive, found a tunnel on 71 of the 79 bridge drives
scanned. Seven of the eight misses were after dark. <a href="https://looselydecoupled.com/find-landmark-dashcam-video-tunnel-light-color/">Finding a bridge in
dashcam video by the color of tunnel
lights</a> is the full account,
including what didn&rsquo;t work.</p>
<figure><img src="https://looselydecoupled.com/triage-dashcam-footage-claude-code/bridge-daylight-vs-tunnel-sodium_hu_addb249c3b37c46e.jpg" alt="Two front-camera frames side by side. On the floating bridge in daylight the picture is blue-gray, with a red/blue ratio of 0.81. Inside the Mount Baker tunnel it is orange-brown under the sodium lamps, with a ratio of 1.37, and the road and cars are still well exposed." width="1600" height="640" loading="lazy" decoding="async"><figcaption>Mean red over mean blue for each whole frame, from the same drive. Plates blurred.</figcaption>
</figure>
<p><strong>What went wrong here.</strong> The first calibration read zero change between every
pair of keyframes, parked or moving. A signal that&rsquo;s flat on a known positive and
a known negative is a broken instrument, not a finding. The cause was ffmpeg
duplicating keyframes back up to 30 fps, so consecutive frames were identical.
<code>-fps_mode passthrough</code> fixed it. That bug is why the prompt asks for a known
positive and a known negative before the scan.</p>
<p>The automatic pick, one tunnel chosen by its position in the drive, was right
about two times in three. The rest landed on other tunnels. The contact sheet is
what turned an unreliable detector into a usable one. It showed one frame per
candidate, taken from the bridge side of the tunnel. From it, I confirmed 68
crossings, couldn&rsquo;t confirm two night drives, and found none on one. The first sheet was a flat grid
of every candidate frame, and it was easy to match a frame to the wrong drive.
One row per drive makes that mistake visible.</p>
<figure><img src="https://looselydecoupled.com/triage-dashcam-footage-claude-code/review-sheet-one-row-per-drive_hu_bceefb337d7a67ac.jpg" alt="Three rows of a review contact sheet, one drive per row, each tile a frame from the bridge side of one tunnel candidate. Each row has a floating-bridge tile among freeway tiles, so the right pick stands out at a glance. The first row’s last three tiles are replaced by gray blocks." width="1600" height="768" loading="lazy" decoding="async"><figcaption>Three rows of a review sheet. Some frames are removed.</figcaption>
</figure>
<p>The worst wrong claim of the whole session came at this stage. Partway through,
the agent reported that the picks for some drives were the wrong tunnel, the Mercer
Island lid instead of the Mount Baker tunnel. The contact sheets had been built
with an output-side <code>-ss</code>, which ffmpeg applied after tiling, so the sheets showed
a different window from the one the picks named. Extracting one frame two ways
exposed it. The picks had been right, and the claim was retracted.</p>
<h2 id="cut-synced-clips-and-render-hyperlapses">Cut synced clips and render hyperlapses<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#cut-synced-clips-and-render-hyperlapses"
     aria-label="Link to section: Cut synced clips and render hyperlapses" data-pagefind-ignore></a>
</h2>
<p>With picks confirmed, paste:</p>
<!-- style-check: off -->
<div class="code-block code-block--wrap" data-copyable data-copy-label="Text">
  <div class="code-block__head"><span class="code-block__lang">Text</span>
  </div>
  <div class="code-block__scroll">
    <div data-copy-source><div class="highlight"><pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">For each confirmed pick in &lt;picks-file&gt;, cut a clip from every camera, from &lt;seconds-before&gt; seconds before the landmark to &lt;seconds-after&gt; seconds after it, without re-encoding. Then render a 30x hyperlapse of every drive for every camera, decoding keyframes only.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">Rules:
</span></span><span class="line"><span class="cl">- Read only on &lt;footage-folder&gt;. Write only under &lt;work-folder&gt;.
</span></span><span class="line"><span class="cl">- A stream copy can start only on a keyframe. Measure each camera&#39;s keyframe timing first; do not assume the cameras share it.
</span></span><span class="line"><span class="cl">- A camera&#39;s first frames in a segment file can depend on frames at the end of the previous file. Cut across segment files in one pass, and check the joins.
</span></span><span class="line"><span class="cl">- After cutting, compare every camera&#39;s frame count, duration, and start_time. Same length isn&#39;t the same as in sync.
</span></span><span class="line"><span class="cl">- Pad a camera that stopped early with black frames, per segment, so that every camera stays the same length as the front camera. Give the black the footage&#39;s pixel format and color range, and check the range of every joined file.
</span></span><span class="line"><span class="cl">- Do not trust folder names or file timestamps for timing. If a camera restarted into another folder, measure its offset from cameras that see the same scene, in more than one window. If that offset doesn&#39;t land on one of the camera&#39;s keyframes, re-encode only that camera, cut at the exact frame, and tell me which files were re-encoded.
</span></span><span class="line"><span class="cl">- Test on one drive first, and include a drive where a camera stopped early.
</span></span><span class="line"><span class="cl">- Before the batch, shift one camera by a second on purpose and confirm that your sync check fails. Then restore it.
</span></span><span class="line"><span class="cl">- In the batch, one bad drive must be logged and skipped, not stop the run. Write each output to a temporary name and rename it when it&#39;s complete, so that a rerun resumes.
</span></span><span class="line"><span class="cl">- Time the test drive and estimate the batch from that.
</span></span><span class="line"><span class="cl">- If the run will outlive this session, start a watcher that tells me when the run ends or stops making progress.
</span></span><span class="line"><span class="cl">- List the drives that were skipped or padded, and anything you couldn&#39;t verify.</span></span></code></pre></div></div>
  </div>
</div><!-- style-check: on -->
<p>Each keyframe becomes one frame at 30 fps, so a hyperlapse plays at 30× speed. A
36-minute drive became 72 s and took 4.4 minutes to render. The hyperlapses of
218 drives, five cameras each, came to 107 GB. <a href="https://looselydecoupled.com/ffmpeg-cut-multi-camera-dashcam-clips-in-sync/">Cut multi-camera dashcam clips
with ffmpeg</a> and <a href="https://looselydecoupled.com/ffmpeg-keyframe-hyperlapse/">make a
hyperlapse from keyframes with ffmpeg</a> cover each
step in full.</p>
<p><strong>What went wrong here.</strong> On the first drive I cut, the side-left camera came out
short. On other drives it&rsquo;s a different camera, or none, so the prompt measures
every camera. The frame-count comparison caught it twice:</p>
<ol>
<li>Its GOPs run across segment files. Cutting each segment separately and joining
them dropped the frames that reference the previous file: the camera came out
1.9 s short, with a jump at the join. One concat-demuxer pass over the source
segments keeps them.</li>
<li>It was still 29 frames short, because its keyframes fall at n + 0.967 s, not on
whole seconds like the other cameras&rsquo;. Landing each camera&rsquo;s in-point exactly
on its own keyframe makes every file start at time 0. An offset camera
starts a few frames early, and setting that clip&rsquo;s In point that many frames
in lines it up with the others.</li>
</ol>
<figure><img src="https://looselydecoupled.com/triage-dashcam-footage-claude-code/synced-clip-five-cameras_hu_321fa648b5047a9c.jpg" alt="The five cameras of a finished synced clip at the same second on the floating bridge. A container truck passing on the left fills the front and side-left views, and the pickup behind appears in the rear and Gear Guard views." width="1600" height="853" loading="lazy" decoding="async"><figcaption>One frame from each camera&rsquo;s cut file, all at the same timestamp. Plates blurred.</figcaption>
</figure>
<p>A third check, on <code>start_time</code>, caught a version that had the right length but
started 0.967 s late, which an editor might ignore.</p>
<p>The hyperlapse test drive caught two more. Two cameras came out at 1,303 frames
against the front camera&rsquo;s 2,170. The first guess was the black filler&rsquo;s pixel
format; the real cause was the rear camera stopping 33 s into a segment. The fix
is to pad every segment to the front camera&rsquo;s length. The filler has a trap of
its own: it has to be full range like the footage, which takes <code>-color_range pc</code>;
Apple&rsquo;s hardware encoder ignores <code>-pix_fmt yuvj420p</code>. A mismatched part doesn&rsquo;t make the stream-copy join fail.
The joined file takes the first part&rsquo;s label, so a camera whose file opens on
filler was labeled limited range throughout: 11 of more than a thousand hyperlapse files, found
and re-rendered later. Across 233 drive folders, six hyperlapse sets needed
black-filled segments.</p>
<p>On that same drive, the rear and side-left cameras had restarted into a sibling
folder whose name was 34 s later. The name was wrong by more than a second.
Correlating motion between the Gear Guard camera and the restarted rear camera,
which both face backward, measured 32.7 s ± 0.1 s across four independent
windows. Front against rear disagreed by 0.9 s between windows, because those two
cameras don&rsquo;t see the same scene. <a href="https://looselydecoupled.com/rivian-dashcam-camera-missing-sibling-folder/">Rivian dashcam camera missing from a
drive</a> covers the restart.</p>
<p>Then the batch failed, and no check caught it. Background jobs in the session
stopped at a time limit, so the run was relaunched detached. A detached process
sends no notice when it ends. It crashed at 10:11 on one of the empty folders the
moves had left behind, and it sat dead until 16:48, when I asked what was next:
more than six and a half hours. Two
fixes came out of it, and both are in the prompt: an empty or bad drive is now a
logged skip, and a long run gets a watcher.</p>
<h2 id="keep-location-out-of-anything-you-publish">Keep location out of anything you publish<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#keep-location-out-of-anything-you-publish"
     aria-label="Link to section: Keep location out of anything you publish" data-pagefind-ignore></a>
</h2>
<p>Paste, for each file you mean to share:</p>
<!-- style-check: off -->
<div class="code-block code-block--wrap" data-copyable data-copy-label="Text">
  <div class="code-block__head"><span class="code-block__lang">Text</span>
  </div>
  <div class="code-block__scroll">
    <div data-copy-source><div class="highlight"><pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">Prepare &lt;export-file&gt; for publishing.
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">Rules:
</span></span><span class="line"><span class="cl">- Treat GPS in any clip as private. List every metadata field that holds a location, with: exiftool -G1 -a &#34;-gps*&#34; &#34;-location*&#34; &lt;file&gt;
</span></span><span class="line"><span class="cl">- First run that command on an original clip from the truck and confirm that it finds the location. Then run it on the export.
</span></span><span class="line"><span class="cl">- If the export carries a location, write a copy without it, run the same command again, and show me the empty result.
</span></span><span class="line"><span class="cl">- List the first and last frames of every clip, and flag any frame from either end of a drive, any license plate, and any face. Do not decide for me what to blur or cut.
</span></span><span class="line"><span class="cl">- Also list creation dates, and any drive date or time in file names. Strip them with -map_metadata -1 and rename the files, unless I say otherwise.
</span></span><span class="line"><span class="cl">- List what this check can&#39;t see, such as a street sign or a house number in the picture.</span></span></code></pre></div></div>
  </div>
</div><!-- style-check: on -->
<p>Run the check on every export rather than reasoning about which files carry a
tag. In this work the answer differed by tool. The raw clips carry it. The
hyperlapse tool I used copies the first source file&rsquo;s metadata, GPS included,
so the set stays traceable; the <a href="https://looselydecoupled.com/ffmpeg-keyframe-hyperlapse/">published hyperlapse
script</a> strips it. Clips cut with the concat demuxer came
out without it, but a plain <code>-c copy</code> of a single file keeps it, and
<code>-map_metadata -1</code> removes it. Show only
public landmarks on screen, and never frames from either end of a drive.</p>
<figure><img src="https://looselydecoupled.com/triage-dashcam-footage-claude-code/interchange-curve-sunset_hu_6d94b69d2acd59a0.jpg" alt="A side-camera fisheye frame on the I-5/I-90 interchange curve at sunset: the low sun flares at the left edge, and the downtown Seattle skyline stands above the ramps and trees." width="1600" height="1280" loading="lazy" decoding="async"><figcaption>A public landmark mid-drive, raw fisheye, which keeps the sun in frame.</figcaption>
</figure>
<div class="admonition admonition--caution" role="note">
  <p class="admonition__label">
    <svg class="icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><use href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#i-octagon-alert"/></svg>
    <span>Caution</span>
  </p>
  <p>A clip from the truck&rsquo;s drive carries a GPS fix for the start of every segment, and
the first segment of a drive usually starts at home. Run the <code>exiftool</code> check on
every file before it leaves your machine.</p>
</div><!-- style-check: end procedure -->
<h2 id="detectors-that-didnt-work">Detectors that didn&rsquo;t work<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#detectors-that-didnt-work"
     aria-label="Link to section: Detectors that didn’t work" data-pagefind-ignore></a>
</h2>
<p>Four ideas failed, and each failure says something about the signal rather than
the threshold:</p>
<ul>
<li><strong>Parked footage.</strong> The hypothesis was that recording while parked padded every
drive. The measurement found 23 GB of 1,760 GB: with Gear Guard turned off, the
truck stops recording seconds after parking. The scan still paid for itself, because its per-second
signals fed every detector after it.</li>
<li><strong>Hard stops.</strong> The detector returned 875 candidates, and all 12 at the top
were turns into parking lots or ordinary stops. Turning changes the picture as
much as braking does.</li>
<li><strong>Traffic jams.</strong> Crawling traffic still changes the picture, because the
neighboring cars move.</li>
<li><strong>Garage boundaries by color.</strong> An overcast sky is as gray as a concrete
ceiling. It worked on a clear day and failed on a cloudy one, which is the case
to test first.</li>
</ul>
<p>One detector worked with a person in the loop. Close passes, read as a
one-sided spike on a side camera while the front camera shows steady driving,
were swamped by overpass shadows until the frame&rsquo;s overall brightness change was
subtracted. Of the top 60, 37 were real passes, and a contact sheet sorted them.
A human pass was cheaper than a better detector.</p>
<h2 id="failures-and-the-checks-that-caught-them">Failures and the checks that caught them<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#failures-and-the-checks-that-caught-them"
     aria-label="Link to section: Failures and the checks that caught them" data-pagefind-ignore></a>
</h2>
<div class="table-wrap" data-scrollable data-label="Table: Failure">
  <table>
    <thead>
      <tr><th scope="col">Failure</th><th scope="col">Caught by</th></tr>
    </thead>
    <tbody>
      <tr><td>No road trips in the footage</td><td>The distance ranking itself</td></tr>
      <tr><td>GPS parser failed on a tag ending in <code>/</code></td><td>An exception on the first run</td></tr>
      <tr><td>Move would copy but never delete</td><td>Inspecting the running process</td></tr>
      <tr><td><code>rm</code> with a possibly unset variable</td><td>Claude Code&rsquo;s command guard</td></tr>
      <tr><td>Every frame difference read zero</td><td>A known parked and a known moving segment</td></tr>
      <tr><td>Automatic tunnel pick wrong about a third of the time</td><td>A one-row-per-drive contact sheet</td></tr>
      <tr><td>Claim that the picks were the wrong tunnel</td><td>Extracting one frame two ways</td></tr>
      <tr><td>One camera short in a synced cut, 1.9 s and then 29 frames</td><td>Frame counts across cameras</td></tr>
      <tr><td>The same camera starting 0.967 s late</td><td>Inspecting <code>start_time</code></td></tr>
      <tr><td>Hyperlapse cameras at 1,303 frames, not 2,170</td><td>Frame counts on a test drive with a stopped camera</td></tr>
      <tr><td>Folder name said 34 s; the offset was 32.7 s</td><td>Motion correlation in four windows</td></tr>
      <tr><td>Detached run dead for more than six and a half hours</td><td>Nothing, until I asked; now a watcher</td></tr>
      <tr><td>Hard stops, jams, and garage color</td><td>A contact sheet, a count, and an overcast day</td></tr>
    </tbody>
  </table>
</div>
<h2 id="each-stage-in-depth">Each stage in depth<a class="heading-anchor" href="https://looselydecoupled.com/triage-dashcam-footage-claude-code/#each-stage-in-depth"
     aria-label="Link to section: Each stage in depth" data-pagefind-ignore></a>
</h2>
<p>Each later part of this series takes one stage and stands on its own:</p>
<ol start="2">
<li><a href="https://looselydecoupled.com/rivian-dashcam-gps-location-ffprobe/">Read the GPS location from Rivian dashcam video with
ffprobe</a>: the tag&rsquo;s spellings, a
parser, and how to strip it.</li>
<li><a href="https://looselydecoupled.com/rivian-dashcam-camera-missing-sibling-folder/">Rivian dashcam camera missing from a drive: it restarted in another
folder</a>: finding the sibling
folder and measuring its real offset.</li>
<li><a href="https://looselydecoupled.com/ffmpeg-cut-multi-camera-dashcam-clips-in-sync/">Cut multi-camera dashcam clips with ffmpeg without re-encoding, and keep them
in sync</a>: snapping each
camera to its own keyframe.</li>
<li><a href="https://looselydecoupled.com/ffmpeg-keyframe-hyperlapse/">Make a hyperlapse from keyframes with ffmpeg without decoding every
frame</a>: the 30× render, the padding, and the
full-range filler.</li>
<li><a href="https://looselydecoupled.com/find-landmark-dashcam-video-tunnel-light-color/">Finding a bridge in dashcam video by the color of tunnel
lights</a>: the detector, the
review sheet, and what didn&rsquo;t work.</li>
</ol>
]]></content:encoded>
    </item>
    
  </channel>
</rss>
