Imported from CloudStone16/ccnotes (.agents/skills/ccnotes-build/SKILL.md). Install upstream with npx skills add CloudStone16/ccnotes --skill ccnotes-build. Copyright stays with the author.

ccnotes-build — pipeline orchestrator

You are the conductor. You never read slide or section content yourself — only file names, counts, and build/PROGRESS.md. Every heavy step is delegated to a subagent that returns a short report. This keeps your context small and the build cheap and fast.

0. Read first (once)

• NOTES_SPEC.md — the whole file.
• The target: a notebook id (or alias) + a deck folder under inbox/. No notebook yet? node tools/new-notebook.mjs "<Subject>" <unitNo> "<Title>" (or bun) — it creates the notebook folder AND the inbox folder, and prints both.

1. Normalize decks (fast, no content reads)

• unzip -o <f>.zip -d <slug>/ for every .zip in the deck folder.
• Ignore PUT-DECKS-HERE.txt and anything that is not .pptx / .pdf / a slides export.
• Order deck files by filename; assign slugs 01-<name>, 02-<name>, ...
• Write content/<subjectSlug>/<id>/build/decks.json: { "decks": [ { "slug": "01-intro", "file": "Intro.pptx", "kind": "pptx", "slides": 0 }, ... ] } (fill slides after extraction).

2. Create the progress tracker

Copy templates/progress.md to content/<subjectSlug>/<id>/build/PROGRESS.md and fill the header + one row per deck. This file is the source of truth for what is done.

• After every subagent returns, tick its box and paste its 1-line result.
• On resume / re-run: read PROGRESS.md first and skip every ticked item.
• Every "Gap in source" or "needs elaboration" note from a subagent goes in the ## Gaps to revisit section — you revisit these before assembling.

3. The schedule (parse decks one by one; overlap everything else)

Rule: decks are extracted strictly one at a time, in order (bounds memory + prevents context rot). But the moment deck k's slides.json exists, its section author can run while deck k+1 is being extracted. Keep at most 3 subagents running at once.

loop over decks in order:
  - Task: ccnotes-slide-extract  deck k          (wait for it — sequential)
  - update decks.json slides count + PROGRESS
  - Task: ccnotes-section-author deck k           (do NOT wait — let it run)
  - continue to deck k+1 extraction in parallel with the running authors
after last deck extracted: wait for all section-author tasks to finish

Collect every author's viz-requests[] and gap notes as they report.

4. Visualizations (parallel within, loop until clean)

• One ccnotes-viz-build subagent, given the full combined viz-requests[] list. If there are many (8+), split into 2 subagents by section range, run in parallel.
• Then one ccnotes-viz-verify -> build/viz-report.json.
• While any verdict != "pass": re-spawn ccnotes-viz-build with only the failing items + the verifier notes, re-verify. Max 3 rounds; then record the failures in PROGRESS ## Gaps to revisit and continue (don't block the whole build on one viz).

5. Assessments — run the 4 sub-skills concurrently (cap of 3 at a time, so 3 then 1, or 2+2)

ccnotes-mcq, ccnotes-flashcards, ccnotes-theory, ccnotes-quickref. Each is told: notebook path, section list, "read the finished sections/*.html and build/*/slides.json for traceability — NOT the raw decks". Each writes its own data/*.json and reports counts.

6. Assemble + validate (sequential)

ccnotes-assemble: builds index.html, isa/index.html, manifest.json, SKILL_SOURCE.md; runs node tools/validate-notebook.mjs <id> (or bun); sets registry status: "ready" only on pass; reports the validator output verbatim. On failure: read the precise error list, dispatch the smallest targeted fix to the one responsible sub-skill (e.g. "flashcards count low" -> re-run ccnotes-flashcards only), tick PROGRESS, re-assemble. Repeat.

7. Doubt skill

On a green validate: node tools/gen-doubt-skill.mjs <id> (or bun). Tick PROGRESS.

8. Report to the user (short)

id + alias · sections · slide coverage (must be 100%) · counts (section MCQs / flashcards / 1-mark / 2-mark / 4-mark / viz) · validator verdict · any unresolved gaps · "run ccnotes open".

Token & speed discipline (do this, it matters)

• Never paste slide, section, JSON, or HTML content into your own context. Work from PROGRESS.md + short reports + ls / counts.
• Every subagent prompt: give paths, not content; demand a report of <= 12 lines (paths + numbers + issues). If a subagent returns pasted content, tell it to compress and resend.
• Each subagent re-reads NOTES_SPEC.md and its own inputs itself — you don't relay spec text.
• Run subagents concurrently up to the cap of 3. Don't serialize things that don't depend on each other (the 4 assessment skills; multiple section authors; viz splits).
• One deck at a time for extraction — never batch decks into one extractor.
• If the build is interrupted, resume from PROGRESS.md; do not rebuild ticked items.