agentleFS
chendanielychendaniely/skills/claude-chats/SKILL.md

claude-chats

Skill1 starsChanged 5 days agoMITView on GitHub

Use when Dan pastes a claude.ai share link (claude.ai/share/…) or a claude.ai chat URL (claude.ai/chat/…) in an Obsidian vault that mirrors his Claude chats (its CLAUDE.md says so), whether alone or with words like "capture this", "add this" or "mirror this"; when he asks to re-sync a chat already mirrored there; when he asks to sync a claude.ai project by name; and when he asks to process the chats a project sync brought in, such as going through his outfits.

2 other projects use a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. About 0 are common, 9 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈9

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the vault's CLAUDE.md paragraph on mirrored Claude chats once per session.

    line 22
  2. Very specific:

    Run every command from the vault root, with S=<this skill's base directory>/scripts/claude_chats.py (for example …

    line 25
  3. Distinctive:

    A whole project, or processing what one brought in: follow references/projects.md.

    line 29
  4. Very specific:

    The vault has no .claude/claude-chats.toml: use capture-conversation.

    line 36
  5. Distinctive:

    2. Fetch, in the main session only: Never hand the browser to a subagent.

    line 44
  6. Very specific:

    Before --apply: if a new note's title suggests a student, a therapy subject or anyone Dan evaluates, stop.

    line 76
  7. Very specific:

    Otherwise: run render --apply, then act on its flags:

    line 80
  8. Distinctive:

    Read the whole note first: long chats change direction late.

    line 96
  9. Very specific:

    Write the summary to a temp file outside the vault, then run uv run "$S" summary <id> --file <tmp>.

    line 97
  10. Distinctive:

    6. Daily notes: Read each note right before you edit it, because other sessions write to these notes too.

    line 115
  11. Very specific:

    Never: write inside the <!-- ea-email:… --> block, and never move or reflow lines that aren't yours.

    line 124
  12. Distinctive:

    If a day's note is missing,: create it from the vault's daily-note template for that date.

    line 126
  13. Distinctive:

    Add actions only, not every open question.

    line 134
  14. Distinctive:

    Leave daily notes that git doesn't track yet for the vault's next checkpoint.

    line 144
  15. Distinctive:

    Use the same steps for "re-sync <link>", or for a link to a chat that's already mirrored:

    line 155

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Claude Chats

Overview

Dan's link is the whole request. Mirror that chat into the vault end to end, then reply in a few lines. This skill's scripts/claude_chats.py fetches, archives and renders; the rest of this file drives it and writes the parts that need judgment: the summary, the daily-note links and the follow-ups.

The link is the yes. Pasting a share link or a chat URL is Dan's request to mirror that chat, and mirroring means saving it, with its photos, from his own claude.ai account into his own archive. That download is the request, not a step beyond it, so don't ask before it: Dan decided this on 2026-11-01 ("i'd rather just give you a share link and you process it in the vault entirely"). The yes covers that one chat and nothing else: not other chats, and not a project's instructions or files.

Read the vault's CLAUDE.md paragraph on mirrored Claude chats once per session. Where it and this skill differ, the vault wins.

Run every command from the vault root, with S=<this skill's base directory>/scripts/claude_chats.py (for example ~/.claude/skills/claude-chats/scripts/claude_chats.py).

A whole project, or processing what one brought in: follow references/projects.md.

When not to use

  • The chat isn't Dan's: the snippet says so (below). Stop, and offer capture-conversation, which captures someone else's share from the page text.
  • The vault has no .claude/claude-chats.toml: use capture-conversation.

Steps, for each link in turn

1. Snippet. From the vault root: uv run "$S" snippet '<link>'. Exit 64 means it isn't a claude.ai share link or chat URL: say so and stop. Pass the output on whole; never edit or retype it.

2. Fetch, in the main session only. Never hand the browser to a subagent.

  • tabs_context_mcp (createIfEmpty), then tabs_create_mcp for a fresh tab. Never use one of Dan's tabs.
  • navigate that tab to https://claude.ai/robots.txt.
  • Google Chrome only, never Brave. Dan keeps Brave for himself, and both have the extension. Run (navigator.userAgentData?.brands || []).map(b => b.brand) with javascript_tool, and go on only if it lists Google Chrome. If it lists Brave, close the tab and get Chrome connected (list_connected_browsers and select_browser, or switch_browser with Dan clicking Connect in Chrome), then start again in a new tab.
  • javascript_tool with the snippet's output.
  • Next link: a new tab. Chrome asks before a page starts a second automatic download.
  • Close each tab only once step 3 has its file: a download Chrome is holding lives in that tab, and closing it loses the download.
The snippet returnsThen
saved, id, photos, photosMissinggo on
held: truestop for this link: "That chat matches a hold rule, so nothing was saved." Never guess at or repeat the title
notYours, or "this account can't open that share link"stop: it isn't in Dan's account (see When not to use)
"not signed in to claude.ai"ask Dan to sign in to claude.ai in Chrome, then run it again
"couldn't read the chat's project"its project's hold rules couldn't be checked, so nothing was saved. Run it again in a new tab; if it fails twice, tell Dan
the tool call times outdon't run it again: the page keeps going and saves the file. Leave the tab open and go to step 3

3. Ingest. uv run "$S" ingest --wait 90.

  • Exit 66 means the file never landed, so Chrome is asking: a Save dialog, or "download multiple files" (Allow). Its "Ask where to save each file" setting is meant to be off.
  • Tell Dan that in one line and run the wait again after he answers.

4. Render. Dry run first: uv run "$S" render.

  • Before --apply: if a new note's title suggests a student, a therapy subject or anyone Dan evaluates, stop. The vault's capture guard keeps that out of sources/, and the chat is still only in the archive, outside the vault. Tell Dan, and offer to add its ID to exclude in the archive's config.toml.
  • Otherwise run render --apply, then act on its flags:
FlagThen
holds <% … %> textthat note isn't written, because Templater would run it. Say so and point to the vault's TODO.md item
⚠ non-additivethe chat was edited on claude.ai. The note now follows the current branch, and the old one stays in the archive. Mention it
missing notesDan deleted that note. Never pass --recreate; mention it
project name unknownthe chat was archived before its project's name was. Syncing that project fixes it; tell Dan
summary may be out of datethe chat has grown: in step 5, add only what's new, with --update
couldn't be renderedthat chat's archived copy has something the renderer can't read (exit 1; the other chats are written). Report its ID and carry on
photos that can't be readthe note says "not downloaded" for it. Mention it

Updates to other chats in the same render are fine; mention them in the reply.

5. Summary. uv run "$S" summaries lists the notes waiting for one.

  • Read the whole note first: long chats change direction late.
  • Write the summary to a temp file outside the vault, then run uv run "$S" summary <id> --file <tmp>.
  • For a grown chat whose summary is already written, write only what's new and add --update. It goes under the summary as a dated ### Update, and the summary itself stays as it is, in case Dan has edited it.

Match the shape of the summaries already in the vault's sources/claude/:

  • An opening paragraph: what the chat was about, when Dan had it and how (by voice or text, with photos or not), and why it came up.
  • Where it landed: the decisions or the plan, as bullets.
  • Considered and dropped: each with why.
  • Still open: unresolved questions and next steps.

Link existing vault notes on the same subject, and only ones that exist. Search for them first, and never search under private/. Take facts from the chat alone and mark anything inferred. Never write <%.

6. Daily notes. Read each note right before you edit it, because other sessions write to these notes too.

  • The day or days Dan had the chat: list them with grep -oE '^### \[[0-9]{4}-[0-9]{2}-[0-9]{2}' <note> | sort -u. Under ## Log in each of those notes, add one line: - [[<stem>]]: a Claude chat (by voice, if it was), HH:MM–HH:MM. <one-line gist>. Mirrored on <today>.
  • Today's note: under ## Log, add the line render --apply printed (- Synced Claude chats: …) and where the follow-ups went.
  • Never write inside the <!-- ea-email:… --> block, and never move or reflow lines that aren't yours.
  • If a day's note is missing, create it from the vault's daily-note template for that date. If the template can't be used, put that day's link in today's line instead and say so.

7. Follow-ups. Each thing the chat leaves Dan to do goes into the vault's TODO.md, in the matching section, as - [ ] **<action>** <detail> (from [[<stem>]]) *(added YYYY-MM-DD)*.

  • Add actions only, not every open question.
  • One with a date, or with someone waiting on it, crosses Dan's YouTrack bar. Propose the ticket in the reply; don't file it.

8. Commit what you wrote: the notes, their photos in ze-files/claude/, any project note, and your TODO.md lines.

  • Other sessions may have uncommitted edits in the same files, so stage only yours. For a shared file, git hash-object -w a copy of HEAD plus your lines, then git update-index --cacheinfo.
  • Leave daily notes that git doesn't track yet for the vault's next checkpoint.

9. Reply, a few lines per chat:

  • [[<stem>]], with a one-line gist, the number of messages and photos;
  • the days it's linked from;
  • what went to TODO.md, or "nothing";
  • anything that stopped or was flagged.

Re-syncing

Use the same steps for "re-sync <link>", or for a link to a chat that's already mirrored:

  • ingest says unchanged: there's nothing to write, unless photos were missing last time.
  • The chat grew: the transcript updates, and step 5 runs with --update.

Common mistakes

MistakeInstead
Asking "save this file?" before fetchingThe link is the yes (Overview)
Reading the code to work out the flow firstThe steps above are the flow. --help lists the flags
Running the snippet again after a timeoutThe page is still saving the file; go to step 3
Two links fetched in one tabA new tab for each link
Running the snippet in whichever browser is connectedCheck it's Google Chrome first (step 2); never Brave
Summarizing from the first screen of a long noteRead it all
Linking notes that might existSearch first; link only what's there
Linking the chat only from today's noteAlso link it from each day Dan had it

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.