askit-build-skill
Creates and improves agentskills.io skills to the Advanced Skill Library Standard. Use when you need to author a new SKILL.md, scaffold a skill directory, or raise an existing skill's conformance and description quality.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- askit-build-skill
- metadata
- { "tier": "universal", "chain": "askit-skill-author, askit-reviewer", "version": "0.1.3", "audience": "beginner" }
- description
- Creates and improves agentskills.io skills to the Advanced Skill Library Standard. Use when you need to author a new SKILL.md, scaffold a skill directory, or raise an existing skill's conformance and description quality.
02 资源文件 9
- 参考文档evals/askit-build-skill-triggering.eval.json2.5 KB
- 参考文档evals/README.md361 B
- 参考文档examples/anti-vague-description.md9.5 KB
- 参考文档examples/golden-1-conventional-commit.md6.2 KB
- 参考文档examples/golden-2-incident-postmortem.md11.8 KB
- 参考文档examples/golden-3-dependency-audit.md12.2 KB
- 参考文档README.md669 B
- 参考文档references/authoring-guide.md13.0 KB
- 参考文档references/skill-craft-rubric.md8.1 KB
03 SKILL.md 正文
askit-build-skill
Purpose
Author conformant skills. Two modes: create scaffolds a new skill that passes Bronze on first run; improve raises an existing skill toward the quality bar. Authoring depth is in references/authoring-guide.md.
When to use
When the user asks to create, scaffold, write, or improve a skill.
create mode
- Brief interview: ask for the skill name (kebab-case), what it does, when to use it, and a few trigger keywords. Skip the interview if these inputs are already provided in context.
- Delegate the drafting to the
askit-skill-authorsubagent (the declared chain) when running in a harness that can dispatch it; otherwise perform the steps below inline. Either way the procedure is the same: createskills/<name>/and copytemplates/SKILL.mdintoskills/<name>/SKILL.md. A repository's own skill template governs over this toolkit's. Before copying, look for one: a path the repository's authoring or contributing docs name, or a file such asdocs/internal/skill-template.md. When one exists, copy it instead, and usetemplates/SKILL.mdonly for what it leaves out. A house template usually encodes rules the repository's own gate enforces, so the generic template would produce a skill that gate rejects. - Fill the frontmatter:
nameequal to the directory, and adescriptionthat states what AND when with concrete keywords (see references/authoring-guide.md for the bar). - Scaffold
references/if the skill needs depth. Samples inexamples/are optional at Bronze - add them only if useful. Do not assume the surrounding plugin anatomy exists - this skill works a la carte. - Emit native manifests for the declared targets: set
agent-targetsin the plugin'slibrary.json(for example["claude", "codex"]), then runnode scripts/generators/gen-manifest.mjs <plugin-root> --write --target=all. To emit for one agent only, pass--target=claudeor--target=codex. The plugin'slibrary.jsonagent-targets(not a CLI flag) declares which targets it requires;gen-manifestwrites.claude-plugin/plugin.jsonand/or.codex-plugin/plugin.jsonfromlibrary.json. - Assess the new skill with
node scripts/evaluate.mjs skills/<name> --json(this is what theaskit-evaluateskill runs), report the result, and iterate until it passes with 0 errors.
improve mode
- Run
node scripts/evaluate.mjs <skill> --jsonand read the report. As in create mode, the rewrite work itself may be delegated toaskit-skill-authoror done inline; the steps below are the procedure either way. - For each finding: a samples warn -> add representative examples; a low description score (U5) -> rewrite the description to clear the bar; an over-budget warning (U7) -> move depth into
references/. For any other finding, read its message and apply the fix it states. - Re-run evaluate to confirm the findings are resolved. That closes phase 1 (conformance).
improve mode, phase 2: the craft review (optional)
Phase 1 proves the skill obeys the rules. Phase 2 asks whether it is any good as a teacher. It is OPTIONAL, it is OFFERED not assumed, and it never edits meaning. Decision and contract: ADR 0037 (the builder craft pass), docs/internal/decisions/0037-builder-craft-pass-and-safe-judgment-partition.md.
- Check eligibility. Phase 2 is offered only when the deterministic gate is already clean. Pass the phase-1 gate result (
{ exitCode, errors, warns }) tophaseTwoEligible()fromscripts/lib/craft-review.mjs. If it returns ineligible, say why and go back to phase 1; the craft pass is never a way around a conformance failure. - Offer it, and stop. Tell the user what the craft review costs (one advisory subagent run, see
docs/reference/token-usage-estimates.md) and what it produces. Run it only on an explicit yes. - Dispatch the reviewer. Delegate to the
askit-reviewersubagent (the declared chain), briefed with references/skill-craft-rubric.md. It scores five dimensions - description and trigger quality, instruction clarity, example depth including golden and anti-example presence, reference structure, token economy - and returns findings in the rubric's contract. If the session has noaskit-revieweragent type, as in fallback mode, where the toolkit is a sibling checkout rather than an installed plugin, brief any capable general-purpose agent instead. Give itagents/askit-reviewer.mdand the rubric, and run its output throughtoReviewAdvisory()as usual. - Render a durable report. Write the findings through
toReviewAdvisory()to a JSON file, then render:node scripts/evaluate.mjs <skill> --report=review --advisory <file.json> --format=md --out <report.md>(use--format=htmlfor the shareable page). The craft review is an artifact, not chat output. It renders BESIDE the verdict and cannot change it. - Partition. Call
partitionCraftFindings(). SAFE is a closed allowlist of three mechanical categories (broken link, formatting, missing bookkeeping frontmatter field), each needing a bounded single-line fix. Everything else, including any category the allowlist does not name, is JUDGMENT. - Offer the SAFE subset only, item by item. Show each SAFE fix as the exact before and after. On explicit consent, call
applySafeFixes(<skill>, safe, { consent: true }). Withoutconsent: trueit writes nothing, and it refuses any finding that is not SAFE even if it is handed one. - Report the JUDGMENT findings, do not act on them. Present each with its recommended change and leave the files untouched. If the user asks for one, that is ordinary improve-mode work (step 2), consciously chosen.
- Re-run the gate (
node scripts/evaluate.mjs <skill> --json) to confirm it is still clean after the apply, and say so.
Scope
Emits for Claude and Codex: library.json.agent-targets declares which targets the plugin requires, and gen-manifest.mjs --target=all|claude|codex generates the matching native manifests. This skill delegates authoring to the askit-skill-author subagent (permitted in agents/_chain-permitted.yaml), which in turn delegates assessment to askit-evaluator, and delegates the optional phase-2 craft review to askit-reviewer (the same subagent askit-evaluate uses for its advisory review layer).
04 同仓库更多 Skills
fx-demo
A fixture skill that exists so the re-derived manifest has a component list to compare. Use only from tests/unit/self-consistency.test.mjs.
fx-demo
A fixture skill that exists so the re-derived manifest has a component list to compare. Use only from tests/unit/self-consistency.test.mjs.
fx-demo
A fixture skill that exists so the re-derived manifest has a component list to compare. Use only from tests/unit/self-consistency.test.mjs.
sf-caller
Calls the worker subagent to do a thing. Use when delegating work to sf-worker.
sf-do-thing
Converts CSV input into a formatted summary. Use when the user asks to summarize or tabulate spreadsheet data.
root-skill-plugin
Summarizes a changelog section into one release-note sentence. Use when the user asks to condense a list of changelog bullets into a single announcement line.