demo2skill

Synthesize a SKILL.md from a recorded demonstration of a task. This is the “Record & Replay” parity piece: the user shows the workflow once (in a recorded shell, or via Chrome’s Recorder), and this skill turns the recording into a skill draft — rather than /skillify, which reads the conversation.

Subscription-safe: all synthesis runs in this Claude session (or a headless claude -p child with ANTHROPIC_API_KEY scrubbed). Never call a paid API.

Modes (one command to remember: /demo2skill)

• /demo2skill record [slug] — start a recording. The recorder wraps script and needs a live terminal, which cannot run inside a tool call. So this mode prints the paste-ready command for the user to run in their own prompt: !~/.claude/scripts/demo-record.sh <slug> (the leading ! runs it in the user’s session). Tell them: run it, do the task, type exit, then come back and run /demo2skill to synthesize. Default slug demo.
• /demo2skill (no args) — synthesize from the newest recording (auto-detect).
• /demo2skill --cli <transcript.log> — synthesize from a specific transcript.
• /demo2skill --browser <export.json> — synthesize from a Chrome DevTools Recorder export.

Inputs

• --cli <transcript.log> (default) — a transcript from ~/.claude/scripts/demo-record.sh.
• --browser <export.json> — a Chrome DevTools Recorder export (Puppeteer/JSON flow).
• No path → auto-detect the newest file in ./.omc/artifacts/demos/ then ~/.claude/demos/.

Workflow — CLI layer (--cli)

1. Clean the transcript: ~/.claude/skills/demo2skill/scripts/clean-transcript.sh <transcript.log> (strips ANSI/control noise + script banners).
2. Extract from the cleaned text: the ordered command sequence, the working directory/branch context, and any obvious decision points (a command that was retried, edited, or followed a failure).
3. Quality gate (from skillify): proceed only if all true — “couldn’t be Googled in 5 min”, “specific to this user/project/workflow”, “took real operational effort”. If it fails, say so and stop — don’t make a slop skill.
4. Ask ≤2 clarifying questions via AskUserQuestion: (a) what are the inputs/variables that change between runs? (b) what is the success criterion? Keep it to two.
5. Synthesize a SKILL.md into ~/.claude/skills/<new-slug>/SKILL.md using the frontmatter shape below. Body = Goal, Inputs, ordered Steps (the extracted commands, generalized with the variables from step 4), Verification, Pitfalls.

Workflow — browser layer (--browser)

Honest constraint: the chrome-devtools MCP drives the browser; it cannot passively log a human’s clicks. The capture bridge is Chrome’s built-in Recorder panel (DevTools → Recorder → record → export as JSON). This skill ingests that export.

1. Parse the Recorder JSON: each steps[] entry has a type (navigate | click | change | keyDown | waitForElement | setViewport) and selectors/url/value. Read title + steps.
2. Translate into a human-readable procedure (one line per meaningful step; collapse setViewport/keyDown noise).
3. Quality gate + ≤2 clarifying questions as above.
4. Synthesize SKILL.md with: the readable procedure, plus an optional Playback section expressing the steps as chrome-devtools MCP calls (navigate_page → click → wait_for) or a Playwright snippet, so the skill can re-run the flow later.

Output frontmatter (required)

---
name: <kebab-slug>
description: >
  <what it does> Triggers: "<phrase>", "<phrase>". Do NOT use for <anti-trigger>.
metadata:
  user-invocable: true
  slash-command: /<slug>
---

description MUST be < 1024 chars and lead with verbs + triggers (per the user’s skill-authoring convention).

Pre-ship checklist (from skill-authoring)

• Description leads with what it does + concrete trigger phrases + an anti-trigger.
• Steps are generalized (variables, not the literal one-off values from the demo).
• A Verification section exists and is runnable.
• No secrets/tokens copied verbatim from the transcript — redact to ${ENV_VAR}.
• Slug directory is ~/.claude/skills/<slug>/ with SKILL.md (exact case).

Related

• /skillify — conversation-based extraction (no recording).
• ~/.claude/scripts/demo-record.sh — produces the CLI transcript this consumes.