skill-drafting

Skill Drafting

Use this skill to turn a vague request for an agent skill into a focused, validated skill package. The skill applies the repo's Spec -> Verifier -> Environment method to skill design, then routes detailed guidance through references only when needed.

The authoritative framework spec — the gold-standard package shape, the end-to-end ingest -> build -> finalize -> validate pipeline, the --verbatim mode, and the in-place maintenance model for already-shipped skills — is `better-skill-framework.md` in the repo root.

Required Workflow

1. Read repo instructions first, especially AGENTS.md, CLAUDE.md, and the three-layer method when present.
2. Interview the user through staged gates before nontrivial writes:

- Gate 1: goal, consumers, scope, out of scope, examples, cost of error. - Gate 2: skill category, positive triggers, paraphrased triggers, negative triggers, overlap boundaries. - Gate 3: package anatomy, reference split, scripts/assets, setup/config, memory-file impact. - Gate 4: quality harness, observation artifact, verifier, forward-test plan. - Gate 5: final implementation plan, files, acceptance criteria, and rollback or preservation rules.

1. Classify the skill using Anthropic's three skill use-case categories (Document & Asset Creation, Workflow Automation, MCP Enhancement) before drafting; if it spans categories, name the primary category and explain why.
2. Draft for progressive disclosure: frontmatter is the trigger surface, SKILL.md is the runtime guide, references are opened on demand, scripts/assets exist only when they change behavior or reliability.
3. Require a Gotchas section for recurring failure modes and a Verification section with observable checks.
4. Build or specify a task-specific quality harness that can distinguish good output from bad output.
5. Give the implementing agent a way to observe work while building, such as tests, logs, screenshots, traces, sample runs, or fixtures.
6. Use independent subagents for nontrivial skill architecture critique and forward-testing when the runtime supports them. If unavailable, run clearly separated self-review passes and say so.
7. Assess whether AGENTS.md or CLAUDE.md need tool, hook, guardrail, or routing updates. Propose changes only; do not edit memory files without explicit user approval.
8. Validate the finished skill before claiming success.

Task Router

Building A Reference Skill From A Corpus

When the skill is a large body of ingested documentation (many source chunks rather than hand-written guidance), build it deterministically with the pipeline instead of authoring files by hand.

1. Ingest live docs into a corpus JSONL with scripts/skill_builder.py ingest <format> (a shared HTML->Markdown converter backs the HTML formats): ingest mdbook (an mdBook print.html), ingest html (crawled HTML pages / a docs bundle), ingest rustdoc (generator-rendered API pages; content selector, dropped section ids, and stripped label are options with rustdoc defaults), ingest pdf (born-digital PDF via pdftotext + qpdf). Each record is {chunk_id, title, source_url, text, tags[, subskill, section]}.
2. Build with scripts/skill_builder.py build: it packs chunks into right-sized per-subject reference files at natural source boundaries and writes references/INDEX.md + references/topics.json + a starter SKILL.md — flat, or a router of sub-skills. Default mode resolves source links/images and normalizes formatting; the `--verbatim` preset turns that off for faithful reproduction of already-clean docs. Pass --verify to audit coverage, link/image residue, and file sizes.
3. Finalize with scripts/skill_builder.py finalize --skill <dir> --meta <meta.json> to write the gold SKILL.md sections + sibling GOTCHA.md (routers: every sub-skill).
4. Validate with scripts/skill_builder.py validate (--package for routers).

See `better-skill-framework.md` for the full pipeline and the gold package spec.

Breaking Up An Oversized Reference Doc

When a skill already has good prose but a few references/*.md files are too big to load efficiently, split them deterministically with scripts/skill_builder.py split — no LLM, no rewriting. Each input doc is parsed at a heading level (default ###, optionally limited to one ## section), CONSECUTIVE same-title blocks are merged into one topic, and each topic becomes its own file (or, with --max-bytes, whole topics are packed up to a size). Filenames are slug(title), prefixed with a source subject token only on cross-file collision. It regenerates references/INDEX.md (replacing just the topic-listing section, preserving the rest) and references/topics.json, optionally remaps a symbols.json's reference_files (--remap-symbols, with --symbols-source-map to map each group to its source doc's topic files — far more reliable than term matching when group keys are normalized labels), and removes the originals (--replace-inputs). The split is VERBATIM by default (use --clean only for noisy source); --verify proves byte-for-byte coverage and INDEX/topics consistency. Drive per-file params with CLI flags or a --legend JSON; see python scripts/skill_builder.py split --help.

Recontextualizing A Verbatim Corpus Before Publishing

When a skill's references/*.md were ingested verbatim from upstream docs, reword them into original prose (identifiers/code/links/numbers/tables preserved exactly) before publishing — the licensing gate. The engine is scripts/builder_components/recontext_core.py; drive it with scripts/skill_builder.py recontext:

integrate -> finish -> reconcile -> promote (roots/owner from a --config JSON or CLI flags; nothing hardcoded). Rewriting subagents must go through the **locked, gated writer** scripts/skill_builder.py recontext-subagent (prepare/show/submit): it derives all paths internally, confines writes to one --work-root, and runs Gate A (identifiers), Gate B (~13-word residue), and Gate C (cruft) before writing — so a PASS is verified. recontext drain generates a Workflow that drives this writer (replacing freehand _rw_/_pkt_ writing). Full details, the gate definitions, and the locked-writer contract are in [references/recontextualization.md`](references/recontextualization.md).

Maintaining An Already-Shipped Skill In Place

When a skill is already built and shipped but its source corpus is gone, do not re-render it. Use scripts/skill_builder.py maintain <skill_dir> to AUDIT gold-standard conformance (oversize files, INDEX/topics drift) read-only; add --apply to split oversize, heading-structured reference files in place and surgically add the new parts to references/INDEX.md + references/topics.json, leaving the bespoke SKILL.md / GOTCHA.md untouched. It is router-aware (per sub-skill) and conservative: a file with no ## / ### sub-headings is reported as an ATOMIC outlier and not auto-split (use --force for a size-based cut between blocks, or note it as an accepted outlier). Use --cross-link to add/refresh a conservative ## See also footer on each reference file (links related topics by distinctive name; already-linked files excluded). Run it against the skill in the store (or via either junction — it is the same copy); both .agents and .claude surfaces reflect the change automatically, so there is no separate mirror to update or parity to diff. Separately, scripts/skill_builder.py lint <skill_dir> is a read-only health check that reports dangling local links and INDEX/topics drift to AI/lint/<skill>.md (maintainer-facing; not shipped inside the skill).

Building A Cross-Skill Master Index

To catalog every skill in a skills root, run scripts/skill_builder.py index <skills_root> --mirror <other_root>. It writes a top-level INDEX.md (skill catalog, entity -> skill map, related skills) to both mirrors. Add --seed-covers to seed or refresh a covers: frontmatter list on each SKILL.md (routers: sub-skill area names; flat: top derived entities); the index prefers covers: and otherwise derives entities from topics.json. This is a discovery and audit surface, not auto-routing — agents still route via each skill's description. The master INDEX.md is public (it lives in the skills area); the lint report under AI/lint/ is maintainer-local and is not.

Gotchas

Recurring failure modes and what to do instead live in the sibling GOTCHA.md.

Verification

For this skill package, run the validator against the package you are editing:

python "$env:DEVROOT\SKILLS\skills\skill-drafting\scripts\skill_builder.py" validate "$env:DEVROOT\SKILLS\skills\skill-drafting"

For long or batched skill builds, check usage after each batch. For Claude, use scripts/claude_usage_check.py to print utilization only. For Codex, use scripts/codex_usage_check.py in read-only mode. By default the Codex agent sandbox blocks direct egress to the ChatGPT usage endpoint, so an in-sandbox agent cannot read Codex usage on its own; the user must run the host-side bridge: ask them to launch the visible scripts/codex_usage_bridge.cmd (or scripts/codex_usage_check.py --run-bridge --login-if-needed), which does the auth preflight by default. Use --login only when the user explicitly asks to authenticate. Normal checker runs auto-detect AI/work/codex-usage-bridge.json or can query --app-server-url ws://127.0.0.1:17342 explicitly. Repo-local Codex mode creates or repairs AI/work/.gitignore before writing local Codex state. If credentials are absent or unavailable, both scripts exit cleanly without leaking token data.

For any skill drafted with this skill, verify:

• •Trigger tests: direct trigger, paraphrased trigger, and negative trigger.
• •Structure tests: SKILL.md, references/INDEX.md, and references/topics.json agree.
• •Workflow test: one representative task can be completed using the skill.
• •Quality test: the harness can reject a bad or incomplete output.
• •Environment test: required tools, permissions, hooks, and memory-file changes are present or explicitly proposed.

Reference Map

Start with references/INDEX.md. Load only the reference that matches the current decision or failure mode.

Inspired by

Includes original content written for this skill, informed by: the repo's three-layer method, Anthropic skill-building guidance, and the system skill-creator skill.