lathe-extend

Lathe — Extend a Tutorial

Add the next part to a stored tutorial. Triggered by /lathe-extend <slug> [guidance…]. The new part must continue that tutorial — its example, its numbers, its voice — not a fresh generic one. Everything about shape, research discipline, and callouts comes from the lathe skill; read it and apply it. The voice is whatever the tutorial was generated in — fetch it from the CLI (see step 1), don't guess or re-pick. This skill only adds what's different about extending: continuity and the extend-start → write → extend-commit handshake.

Protocol

1. Absorb the existing tutorial. Read every part in ~/.lathe/tutorials/<slug>/ (part-NN.md). You need the load-bearing context before you write a word:

   • The controlling example and the fixed numbers (sample rate, page size, buffer size) — reuse them exactly; don't invent new ones.
   • The controlling metaphor, if any, and whether a prior part already retired it (if retired, don't resurrect it).
   • The voice. Fetch the tutorial's recorded voice and apply it for all tonal choices:

     lathe voice show --tutorial <slug>
     

     This prints the wrapped voice spec (guardrail preamble + tone guidance). It resolves the voice recorded on the tutorial's metadata, falling back to the configured default for tutorials stored before voices existed. Don't re-pick or drift the voice; structure and substance still defer to the lathe skill, which wins on any conflict with the voice.
   • The level of the reader.
   • The repo and pinned tool versions in metadata.json (repo, repo_branch, tools) — the new part inherits them. Write against the same versions; don't silently bump to a newer toolchain. (If the reader explicitly wants to move the tutorial to a new version, that's a heads-up that it may be time for a fresh tutorial, not a quiet drift — flag it rather than re-pinning here.)
   • Where the previous part's "What's next" pointed — that's your mandate for this part unless the user's guidance redirects it.

2. Research first. Same discipline as the lathe skill's "Research first" step: actually open 3–8 authoritative sources for whatever this new part introduces, take notes with URLs beside the load-bearing facts, and ground or [!UNVERIFIED]-flag every load-bearing claim. New material gets the same scrutiny as Part 1 did. No web access? Say so in one line and write conservatively, flagging the load-bearing unknowns.

3. Reserve the part. Run:

   lathe extend-start <slug>
   

   It prints a single filename (e.g. part-02.md) and sets status extending. Capture that exact filename. If it errors with "already extending or verifying", stop and tell the user — something is mid-flight; don't force it.

4. Write the part to ~/.lathe/tutorials/<slug>/<printed-filename>.

   • This is the one allowed content write. The skill writes the part markdown directly into the tutorial dir at the reserved path — that is required and deliberate, because lathe extend-commit os.Stats the file there before registering it. (Future reader: don't "fix" this by routing it through a CLI command. There is no such command; the binary owns metadata, the skill writes the part body.)
   • Follow the full Tutorial shape from the lathe skill — hook, ## What you'll build, prerequisites, specific section titles, ## Checkpoint, ## What's next, ## Exercises, ## Sources. Each part stands alone and ends with a Checkpoint.
   • Because this is Part N ≥ 2, open with a > [!RECALL] spaced-retrieval beat: one question on a load-bearing concept from a prior part, phrased so the reader must reconstruct the answer (not just recognize it). See the lathe skill's recall before/after.
   • One file, this part only. Don't write index.md, don't touch earlier parts, don't write more than one part.

5. Commit the part. Run:

   lathe extend-commit <slug> <printed-filename> [--source <url> …] --model "<model you are running as>"
   

   Pass --source once for each source you newly consulted for this part — it folds them into the tutorial's research trail (de-duped). Pass --model with the model you are running as (e.g. --model "Claude Opus 4.8") so the reading-page byline reflects who wrote the latest part (last-writer-wins; omit it only to leave the existing label untouched). This registers the part, clears the pending marker, and resets status to unverified.

6. Tell the user it's added: how to view it (lathe serve), and that verification is opt-in (/lathe-verify <slug> — the "Verify this tutorial" button hands you that command). Then stay in session for follow-ups.

Boundaries

• The only durable-state writes are lathe extend-start and lathe extend-commit. Never edit metadata.json directly.
• Writing the reserved part-content file into the tutorial dir is the sole content write — and it's required (step 4).
• Don't verify, don't write index.md, don't write multiple parts, don't edit existing parts.

Stay in session

You're still their expert guide. Stay available for "make this part harder", "why did we structure it this way", "how'd I do on the checkpoint".