skill-creator

Skill Creator — grill → write

Turn one daily task into one Task Skill. Two phases, same session. No scaffolding for later.

Phase 1 — grill

Run the grilling tree on the daily task. Frontier is every decision with prerequisites settled; ask the whole frontier in one round, with recommended answer, then wait.

Cover in order (keep general, no STT/dossier-specific jargon):

1. YAGNI — does this need a skill or does officecli + copy/manual cover it?
2. I/O — one concrete input (folder/file, e.g. 11. MS BỘ.../1. Hồ sơ thầu:1 as example) and one concrete output (.docx/.xlsx/.pptx/.pdf + path) — use real paths the user can ls.
3. Branches — distinct cases the skill handles (missing file, empty table row, different supplier/template). One trigger phrase per branch.
4. Leading word — shared word in prompts + code + docs that fires the skill.
5. Reference/Template — if L3 Fidelity is required, which real file is the Reference (full Format to clone) and which fields become {{placeholder}} in the Template. Read the Reference via officecli read + unzip check, list candidate placeholders (w:t/a:t/v values) for user to confirm. If no Reference exists, decide whether to create a Template from scratch (markdown → docx then inject placeholders) or to treat the task as L1 (markdown-only).

Completion: every branch has a named trigger phrase and I/O is concrete (real input listed, real output officecli clone → substitute → accept path for L3, or officecli create for L1). For L3 tasks a Reference path and placeholder map exist. No dossier/STT wording here — task-agnostic.

Phase 2 — write

Write per writing-for-agents ladder (SKILL-MECHANICS.md:10 for frontmatter):

1. skills/<kebab-name>/SKILL.md — frontmatter name + description (pointer: front-loaded leading word, one trigger per branch, no synonyms). Body: steps in order, each ends on a checkable completion criterion. Demand is exhaustive. For L3 tasks, steps are clone Reference → substitute {{placeholders}} → accept, not create from markdown.
2. Disclose reference behind a pointer when only some branches need it: references/template.md, references/mapping.md, or references/grill-questions.md. Inline what every branch needs; disclose what only some need. Put task-specific tables (e.g. dossier STT→file mapping) in references, not in this file. For L3, include references/template.docx (or .xlsx/.pptx) as the real Template ZIP with {{placeholder}} inside w:t/a:t/v, plus references/placeholder-map.md listing placeholder → source field.
3. Keep single source of truth; do not restate package.json scripts or officecli help.

Scaffold is 2-4 files (SKILL.md + references/*.md + optional scripts/validate.mjs). No scripts/ until a branch needs code or validate needs deep check.

Validate — 4 gates (run in Isolated Runtime, generic)

1. Name & frontmatter — name kebab-case matches folder, description front-loads leading word + one trigger per branch, no synonyms per writing-for-agents:12. Completion: head -n 5 skills/<name>/SKILL.md passes.
2. Checklist coverage — every branch/item the task requires has an entry in references/*.md (e.g. 22 STT for dossier per Procurement Dossier:1 would be grep "| STT" 22 there, but for non-dossier tasks count the task's own rows). Completion: grep "|" counts match spec, no orphan.
3. Office hygiene — officecli read on each template returns markdown without PK spill. L1: officecli create → accept; L3: officecli clone → substitute → accept then unzip -p word/document.xml | grep -c "<w:tr" ≥ body rows, officecli verify-l3 PASS, file type matches original (.doc stays Composite via cp, .docx stays Word 2007+). Completion: bun tests/isolated-workspace/scripts/test-*.mjs:1 style PASS.
4. Tracer Bullet — bun tests/isolated-workspace/scripts/tracer.mjs:1 for docx|xlsx|pptx|pdf PASS, report.md:1 shows 4 PASS. No global install.

Skipped: deep XML font/header check — add soffice --headless --convert-to pdf visual diff when w:tr counts diverge >20%.

See references/checklist.md (grill) and references/skill-template.md (SKILL shell).