session-summary

session-summary

A skill for capturing the current Claude session into a git repo via the scribe MCP server. The MCP runs as the user (host-side), so it has full filesystem and git credentials — works from Cowork (sandboxed) or Claude Code.

When invoked, follow this procedure.

1. Confirm scribe is reachable and configured

Call mcp__scribe__list_repos. Three possible states:

• No repos configured (empty repos map) → first-run flow. Go to step 2.
• One or more repos configured, current is set → ready. Go to step 3.
• Configured but current is null or invalid → ask the user which repo to use, then mcp__scribe__switch_repo. Then go to step 3.

If the call itself fails ("tool not found"), the scribe MCP isn't registered in this session — tell the user to install the plugin (/plugin install scribe@claude-marketplace) and restart their Claude session, then stop.

2. First-run flow (only if no repos are configured)

The user has scribe installed but hasn't pointed it at a notes repo yet. Ask them:

1. "Where should I save your notes? Paste the absolute path to your cloned notes repo." (Example: /Users/larry/Dev/ground-zero.) The path must be a git repo on this machine — not a URL. If they only have a URL, tell them to git clone it first and come back with the local path.

2. "What short name should I use to refer to this repo?" — used as the name for add_repo and as the identifier when they later have multiple. Suggest one based on the directory's basename. Lowercase, hyphens only. Example: ground-zero.

3. "What user slug should your notes go under?" — defaults to slugified git config user.name. Show the suggestion and let them confirm or override. The slug becomes a top-level subdirectory in the repo (<repo>/<user>/...).

Once you have all three, call mcp__scribe__add_repo(name=..., path=..., user=...). It auto-sets current if there was no current. Confirm with one line: "Configured <name> at <path>, notes will go under <user>/."

Optionally offer: "Want me to verify push access works? I'll run scribe's verify_credentials check." — a good idea on first install.

3. Review the session

Look at what happened in this session:

• Conversation with the user (instructions, decisions, observations)
• Connector tool results (Gmail, Calendar, Drive, etc.)
• Any uploaded files: images, audio, video, PDFs, docs

4. Choose the destination folder

The scribe MCP scopes everything under <user>/<folder>/... automatically — pass folder without the user prefix.

The folder structure within the user's namespace is consumer-defined. The scribe plugin doesn't impose a specific layout. Look at:

• The repo's existing structure (run a quick ls via bash if needed) — match what's already there.
• The repo's CLAUDE.md if it has one — many notes repos document their conventions there.

Common patterns you'll see in notes repos that use scribe:

• clients/<client-name>/ — per-client subfolders for client interactions
• contacts/<category>/<name>/ — per-contact subfolders (vendors, partners, etc.)
• workflows/ — process documentation, recurring patterns
• communications/ — daily rollups (the MCP creates these automatically)
• monday/, linear/, etc. — per-tool capture
• shared/ or team/ — collaborative docs accessible to everyone in the repo
• prds/, ideation/, conventions/ — cross-cutting team docs

If the user's intent is ambiguous between two folders, ask once. Otherwise pick one and proceed.

5. Identify attached media

For every uploaded image / audio / video / PDF / doc in this session, find its host-readable absolute path. Cowork uploads typically live at:

/Users/<user>/Library/Application Support/Claude/local-agent-mode-sessions/<...>/uploads/<filename>

Pass each as { "source_path": "<abs-path>", "descriptor": "<short-label>" }. Do NOT transcribe audio or video unless the user explicitly asks — save the file and reference it in the note.

6. Build the note body

## Summary
<what happened, in the user's voice where possible>

## Decisions
<what was decided and why>

## Next Steps
<who owes whom what, by when>

## Source Material
<links: connector references (Gmail thread id, Drive URL, calendar event), context links>
<relative links to any copied media, e.g. ![photo](media/2026-05-02-walkthrough-photo.jpg)>

Do NOT invent details. If a section wasn't covered in this session, leave it terse or omit it rather than guess.

7. Save via the MCP

Call mcp__scribe__save_session once with everything:

{
  "folder": "clients/jane-doe",
  "slug": "offer-accepted",
  "frontmatter": {
    "date": "2026-05-02",
    "participants": ["Larry", "Jane Doe"],
    "transaction": "123 Main St",
    "tags": ["buyer", "offer"]
  },
  "body": "## Summary\n...\n",
  "media": [
    { "source_path": "/Users/.../uploads/inspection.jpg", "descriptor": "inspection-photo" }
  ],
  "repo": "ground-zero"
}

The repo field is optional — omit to use the current default. Pass it explicitly only when the user said "save this to my X repo" and X is one of their configured repos (verify against list_repos if unsure).

The MCP will:

1. Write <user>/<folder>/<date>-<slug>.md
2. Copy each media file into <user>/<folder>/media/ with canonical names
3. Append to <user>/communications/<date>-rollup.md
4. Update <user>/INDEX.md
5. git pull --rebase --autostash, stage, commit, and push

8. Report back

Tell the user, in one or two lines:

• Relative path of the saved note
• Number of media files copied (if any)
• Whether commit + push succeeded (read git.committed, git.pushed, git.error from the response)

If git.pulled is false (rebase conflict), surface the error verbatim — do NOT auto-resolve. If git.pushed is false but git.committed is true, tell the user the commit is local and push can be retried.

If multiple repos are configured, mention which repo the note went to.