Capture Desktop Session

The device-side mirror of scripts/check_desktop_sync.py. That script governs device-to-cloud sync. This skill governs cloud-to-device capture. claude.ai has no API, so capture is paste-driven: Ane pastes a finished Desktop session and this skill routes its content to the right places on disk.

Step 1 — Get the content

Ask Ane to paste the Desktop conversation, or accept a file path if she gives one. If the paste is long, write it to a temp file first to avoid truncation.

Step 2 — Stage the transcript

Run the helper to stage a verbatim, deduplicated transcript:

python scripts/capture_desktop.py stage --input <tempfile> --title "<short title>"

• Exit 0: prints the transcript path. Continue.
• Exit 3: [DUP] already captured — this session was captured before. Tell Ane and stop unless she confirms she wants to re-route it anyway.

Transcripts are local-only (gitignored, OneDrive-backed). The staged transcript IS the "full log" bucket — no further action for that bucket.

Step 3 — Classify the content

Read the staged transcript. Sort its substantive content into four buckets:

1. Deliverables — documents, analyses, tables, drafts Ane produced.
2. Decisions / rules / lessons — locked decisions, framework calls, guidance on how the system should work, feedback on your behaviour. Capture successes as well as corrections: an approach the transcript shows Ane praised, reused, or locked is a success worth banking as a feedback memory (with Why: and How to apply:) so /improve-system trends confirmed successes, not only corrections.
3. Project-knowledge changes — edits that belong in canonical sources (mel_wiki/ pages, ~/.claude/CLAUDE.md).
4. Full log — already staged in Step 2. No action.

If a bucket is empty, say so. Do not invent content to fill it.

Step 4 — Route each bucket

Deliverables (auto-file). Propose a path under the work folder, save with the Write tool, and name it in the summary. Apply mel_wiki/wiki/concepts/edit-preservation-protocol.md when target file exists.

Decisions / rules / lessons (confirm each). For each item, propose the target — a memory file under the memory dir, a wiki entry, or a feedback memory — and show the exact content you will write. Get a yes/no from Ane before writing. Apply mel_wiki/wiki/concepts/edit-preservation-protocol.md when target file exists.

Project-knowledge changes (confirm each). For each item, propose the exact edit to the canonical source and get a yes/no. Apply the change with the Edit tool, scope-bounded. Apply mel_wiki/wiki/concepts/edit-preservation-protocol.md when target file exists.

• For mel_wiki/ edits: stage for the work-folder commit in Step 5.
• For ~/.claude/CLAUDE.md edits: after applying, commit cross-repo by running bash ~/.claude-config-clone/scripts/sync-from-local.sh --auto "capture-desktop: <one-line reason>". This copies live CLAUDE.md into the claude-config clone, reconciles, commits, and pushes. It exits 0 silently if nothing changed.

Step 5 — Close the loop and commit

If any canonical source was edited (wiki or CLAUDE.md), run the sync check so the re-upload flag for the Desktop project fires:

python scripts/check_desktop_sync.py

Report any [DRIFT] it surfaces. Then commit the work-folder changes (deliverables and mel_wiki/ edits only — transcripts are gitignored). Stage by name, never git add -A:

git add <deliverable paths> <mel_wiki paths>
git commit -m "feat(capture): desktop session <slug> — <what landed>"
git push

Step 6 — Report

Deliver a tight summary:

CAPTURE-DESKTOP — <date>
  Transcript: desktop-captures/<slug>/transcript.md (local)
  Deliverables: <list or none>
  Decisions/rules written: <list or none>
  Canonical edits: <list or none> (CLAUDE.md pushed to claude-config: yes/no)
  Sync check: <in sync / N drift flagged>
  Commit: <short-SHA> pushed / nothing to commit