Standing Mandates
- NEVER write the book, or any chapter, in one pass from the request, and NEVER start stage 2 before the user approves
toc.md. Reader level, order, and scope are decided by the TOC; a one-pass draft hardens them before anyone chose. - NEVER state a version, default value, config name, limit, or benchmark without a source. Write
[확인 필요: <fact>]instead. One wrong default discredits a reference book, and a non-major cannot catch it. - ALWAYS close every analogy to the anchor (e.g. "MySQL에서는…") with where it breaks. A non-major carries an analogy past the point it holds; the break line is what stops them.
- NEVER let the agent that drafted a chapter review it. Review is a separate agent; its findings are applied, not argued.
- ALWAYS stop the review↔revise loop on distinct 🔴🟡 = 0 or after three rounds, and say which — a fourth round is polishing.
- NEVER run
write:like-mewithout ≥ 2 samples the user wrote alone. like-me is the user's voice, not "like other books"; the default copyedit isreferences/house-style.md. - ALWAYS invoke each routed skill by name, visibly (
references/skill-routing.md): without its SME skill a chapter repeats what a general model guesses. Not installed → read itsSKILL.mdfrom the repo, follow it, and say so.
Goal: a non-major reads chapter N knowing only chapters 1..N-1 and the anchor; every fact is sourced or marked; the book moves them from the brief's starting performance to its terminal performance with observable practice evidence; the closing status recounts to the files on disk.
Tech Book
Writes a book as files under tmp/books/<slug>/, ending in book.pdf; outside this repo, if tmp/ isn't gitignored,
tell the user before writing. Resume stages 0–2 from the highest on disk, a chapter from its latest file, a check or
stage-6 step only if its file lacks counts: (tear down its labs by name; same pinned inputs).
Not for: a single document, design doc, or blog post (write:prd, write:adr, write:plans); reviewing existing text
(write:writer-verification); a voice rewrite (write:like-me); a spec from a conversation (write:spec).
Process
- brief.md. Reader level, anchor (what the reader knows), scope in/out, register, sources; starting performance,
one observable terminal performance, the book's throughline, and a capstone with success criteria. One
think:grillround with a recommended answer per question; unanswered →[확인 필요: … / 추천: …]in the brief. The first turn ends here. - toc.md.
think:untangle-thoughts(Outline). One block per chapter in thereferences/house-style.mdformat: goal,requires:(earlier chapters only, order checked with think-tool),entry, observableexit, plausiblemisconception, cumulativeexample-step, learningevidence,length:budget, anchor candidates, andsme:(fromreferences/skill-routing.md). Group chapters into parts only when each marks a capability threshold; trace every chapter exit to the terminal performance and capstone. Sweep it first: every unsourced version, default, limit gets[확인 필요]; list what you checked. Present it with onethink:grillround. Stop: no stage 2 until the user approves toc.md. - glossary.md + example-ledger.md + concepts/NN.md. Once per book; sequential-thinking clusters, think-tool picks
analogies. Per concept: definition, prerequisite, analogy + where it breaks, causal working model, misconception it
corrects, one worked example, one boundary/counterexample, the transfer task and its expected evidence,
[확인 필요]items.knowledge:base-builder/knowledge:queryonly when the user supplied sources. - draft/NN.md.
agents:dispatching-parallel-agents, one agent per chapter, itssme:skill mounted. No worktree (tmp/is gitignored: a worktree strands the file); each writes only its own absolute path intmp/books/<slug>/. Follow the explanatory sequence inreferences/learning-design.md; every non-trivial chapter contains a reasoned worked example, the six elements ofreferences/teaching-elements.md(once each, readable, withinlength:), verification, and a compact answer key;### 장 끝 상태closes the chapter body, before답안과 정리. - review/NN-rK.md. Four concurrent checks, never by the drafter, each to
review/NN-rK-<check>.md, merged per house-style "Review file". Fact: official docs (or user sources) first, cited by URL, SME skill second — an SME-vs-docs conflict is 🔴, docs win; unreachable → web.archive.org copy (URL + capture date), else[확인 필요]; labs by the printed method from N-1'srebuild-end-state.shin an empty dir (references/lab-recipe.md), printed ≠ observed is 🔴.cognition:epistemic-reasoner: analogies and absolute claims.write:writer-verificationreview mode, genredoc, its passes inline in the one check agent.../prd/agents/reader-agent.md, blind (1..N-1 by grep, N cut before답안과 정리): each learning goal at its highest claimed level, at most 5 tests per round, taught, exercised, assessed; each teaching element present. Each check file pins N-1 by path + sha256; N's stopping round (zero or cap) counts only when pinned tofinal/N-1.md(learning-design). - Revise draft/NN.md. A fresh reviser closes every 🔴🟡, text first, then changed labs in one script, replacing
before adding (≤ 15% growth per round, required elements exempt; house-style); round 3 closes by cutting. Back to 4; from round 2 only
the checks that flagged 🔴🟡 re-run (🟢-only: no), reading the diff, but a lab, code, or output change re-runs the fact
check's whole lab chain. At the cap one closure pass, re-checked by a fresh agent in
review/NN-closure.md, fixes or cuts each known-false sentence — not a round; open items go tostatus-NN.md## 남은 항목, never final/. - final/NN.md. 6a per chapter: house-style copyedit, one
write:writer-verificationpass not by the editor, terms intoglossary-delta-NN.md,write:like-meonly with user samples,write:diagramfor figures (SVG+PNG); an agent that neither drafted nor edited re-checks the draft→final diff, itself classing hunks form or meaning (answer key, sources = meaning; rejected → revert orstatus-NN.md). A 6a 🟡: known-false → fixed by the editor and re-checked in that diff; otherwisestatus-NN.md. 6b, after every final: merge deltas, then the book-wide learning pass (references/learning-design.md), everyrebuild-end-state.shreplayed in order. - book.pdf.
node scripts/render-pdf.mjs tmp/books/<slug>— toc.md order, NanumGothic fonts ofassets/fonts/(page, type, figures:references/house-style.md). Exit 1: no browser (book.htmlkept) — ask forCHROME=<path>. Exit 3: PDF written but a chapter or image is missing, or an image is not on its own line — report them; a partial book, never done. Never swap in another font: the page is set to Nanum metrics, and only these ship licensed.
Done: render exit 0, then completion:verification-before-completion with an isolated verifier against the Goal line
and book.pdf (exists, ≥ 1 page, every toc chapter rendered). Parallelism: one agent per book for stages 0–1, stage 2
once per book, 3–6a per chapter, then 6b once per book; no two agents write the same file.
Output Template
Next: <the one decision for the user — e.g. approve toc.md, answer grill items 2·4>
Book: <title> · tmp/books/<slug>/ · stage <n> of 0–7
Chapters <k>: final <a> · in loop <b> (round r/3) · stopped at cap <c>
Review open: 🔴 x · 🟡 y · 🟢 z | learning goals evidenced <e>/<g> · transfer tasks passed <p>/<t> | analogies tested <n>: hold <h> · break stated <b>
Unverified: <m> × [확인 필요] — <list>
Skills run: <plugin:skill × stage>
Before stage 3 the Chapters and Review lines read —. Counts are recounted from the files at report time; a count not
re-measured since the last change names its round, e.g. (r3). Do NOT paste chapters into chat — they are in the files.
What Claude Does / What You Do
| Claude | You |
|---|---|
| Drafts the brief and the TOC, routes each stage to its skill, runs the review loop | Approve toc.md; answer the grill items |
Marks every unsourced fact [확인 필요] |
Confirm those facts, or point to a source |
| Reports the stage it stopped at and why | Decide the leftover 🟢; give writing samples if you want like-me |
Related Skills
references/learning-design.md— book arc, chapter teaching sequence, examples, practice, and evidence;references/teaching-elements.md— the six required teaching elements and their shapes;references/skill-routing.md— which skill runs at which stage, with what input;write:prd/write:adr— a single document.write:writer-verification— review passes, stages 4 and 6;write:like-me— the user's voice, stage 6, with samples.