build-skills（培养技能）

Build Skills

Build or revise Claude skills from evidence, not instinct.

Trigger boundary

Use this skill when the task is about:

• creating a new Claude skill from scratch
• substantially redesigning an existing skill
• merging patterns from multiple skills or source corpora
• tightening a skill's triggers, workflow, routing, or guardrails without drifting from repo conventions

Do not use this skill for:

• tiny wording or formatting cleanups where outside evidence would not change the result
• general repo documentation work that is not about a Claude skill
• copying a remote skill into the workspace with minor renaming

Non-negotiable rules

1. Workspace first. Inspect the current workspace before searching remotely or drafting anything.
2. Research before synthesis for non-trivial work. New skills, major redesigns, multi-source merges, thin local evidence, or explicit comparison requests require remote research.
3. The comparison table is the proof of research. If non-trivial research happened, the comparison table must appear in output before synthesis.
4. Comparison before combination. Build the comparison table before proposing the design.
5. Original, repo-fit output only. Distill patterns; do not rename-clone a source skill.
6. Keep progressive disclosure clean. Put trigger logic in frontmatter, workflow and decisions in SKILL.md, and bulky detail in references/.
7. Test before shipping. Run trigger tests and at least one functional test before declaring done.

Available scripts

Scripts are resolved relative to the skill directory root. Use these instead of inventing ad-hoc shell pipelines:

• scripts/skill-dl — non-interactive wrapper for skill-dl; prefers the globally installed CLI, and on macOS arm64 can fall back to the bundled scripts/skill-dl-darwin-arm64 binary. Run bash scripts/skill-dl --help.
• scripts/skill-research.sh — end-to-end discovery, download, and corpus inspection helper built on top of scripts/skill-dl. Run bash scripts/skill-research.sh --help.
• scripts/skill-dl-darwin-arm64 — bundled macOS Apple Silicon skill-dl binary. Invoke it via bash scripts/skill-dl... rather than calling it directly.

Artifact output

Intermediate artifacts (workspace scan, source shortlist, comparison table, success criteria) appear in conversation output as they are produced — show each one at the step that generates it. Persistent artifacts (final SKILL.md, reference files) live together in the draft skill directory. Default that directory to skills/<skill-name>/; if that path is not writable yet, stage the whole skill folder elsewhere and move it into skills/<skill-name>/ before declaring done.

Required workflow

1. Classify the job

• Local-only path: a small cleanup where triggers, structure, and core method stay intact. Examples: fixing typos, reordering sections, tightening one trigger phrase, updating a version number.
• Full research path: a new skill, substantial redesign, multi-source merge, thin or stale workspace, or an explicit request for outside comparison. Examples: adding a new workflow branch, rewriting the decision tree, changing the skill's category, merging two skills, or any change that alters what the skill does (not just how it reads).

Rule of thumb: if the change would alter the comparison table you'd build for this skill, it's the full research path.

2. Classify the skill type

Before designing, identify the skill's category:

Read references/patterns/workflow-patterns.md to choose the right structural pattern. If the skill enhances an MCP, also read references/patterns/mcp-enhancement.md.

3. Capture local evidence first

• Produce a tree-style inventory of the target workspace.
• If you are revising an existing skill, read the current SKILL.md fully — trigger boundary, workflow, decision rules, output contract, guardrails.
• If you are creating a new skill, read the closest 1-2 local skills plus the naming/format references before drafting so you inherit repo conventions instead of inventing them.
• Read every file in references/ that is relevant to the task: open each file, read the contents, and note what it covers and where it has gaps.
• Note the reference file naming scheme, nesting depth, and whether routing in SKILL.md matches what actually exists on disk.
• Note repo conventions, existing reference coverage, and obvious gaps before widening scope.

Tip: Steps 3–4a are easy to overload. Use the reference routing table to load only 3–5 relevant files per step, not all 22. See references/authoring/reference-file-structure.md for loading discipline guidelines.

4. Run remote research when the job is non-trivial

Only execute this step if step 1 classified the job as Full research path. Skip to step 7 for local-only work (Steps 4–6 produce research artifacts that local-only changes don’t need).

Tip: Downloaded skills frequently violate this skill’s own quality standards (line counts >1000, templates inline, no references). Treat quality problems as signal for your “avoid” column—see references/research/source-verification.md.

• Read references/research-workflow.md for the complete research protocol.
• Verify tool availability with bash scripts/skill-dl --where or skill-dl --version. If skill-dl is missing globally, install it with sudo -v; curl -fsSL https://raw.githubusercontent.com/yigitkonur/cli-skill-downloader/main/install.sh | sudo bash. If installation is not possible, use the skills-as-context-search-skills and skills-as-context-get-skill-details tools (requires the skills.sh MCP server), or search GitHub manually for repositories containing SKILL.md files.
• Use bash scripts/skill-dl search with 3–20 space-separated keywords to discover candidates: bash scripts/skill-dl search mcp server typescript sdk --top 20. It outputs a prioritized markdown table to stdout.
• Use bash scripts/skill-dl to download selected candidates, or run bash scripts/skill-research.sh "keyword1,keyword2,keyword3" for end-to-end parallel discovery and download in one command.
• See references/remote-sources.md for more skill-dl usage patterns and download options.
• Prefer a few diverse, relevant sources over many near-duplicates.
• Show the source shortlist inline in conversation output — list what was downloaded, what made the shortlist, and why — before moving on to Step 4a.

4a. Read the downloaded corpus thoroughly

Do not skip from download to comparison. Reading is a distinct, mandatory step.

For each downloaded skill that made the shortlist:

• Read its SKILL.md fully — not just headings. Understand its trigger boundary, workflow structure, decision rules, and output contract.
• Tree its references/ directory — see what files exist, how they are named, and how deeply they are nested. This reveals the skill's structural philosophy.
• Read the 2–3 most relevant reference files in full — pick based on filenames and the skill's routing logic.
• Check scripts/ if present — scripts surface automation patterns and validation logic that prose cannot fully convey.
• Capture notes per skill (a heading per skill with bullets mapping to comparison table columns): overall structure, workflow style, reference organization, what it does well, what it does poorly, and 1–2 patterns worth inheriting (with exact relative paths).

Downloaded sources may not follow the quality standards this skill enforces (e.g., SKILL.md over 500 lines, missing reference routing). Note these issues — they inform the "avoid" column of your comparison table.

The notes you capture here are the raw material for the comparison table in step 5. If you skip reading, the comparison table will be fabricated from memory rather than evidence.

5. Compare before synthesizing

• Using your notes from Step 4a, build a markdown comparison table with at least: Source, Focus, Strengths, Gaps, Relevant paths or sections, and Inherit / avoid.
• Every row ends with a decision, not just an observation.

6. Define success criteria

Before drafting, write down what success looks like:

• Trigger accuracy: 90%+ of relevant queries should fire the skill; 0% of unrelated queries
• Execution quality: workflow completes without user correction
• Token efficiency: SKILL.md under 500 lines, references loaded only when needed

7. Synthesize the repo-fit result

• Rewrite for the current repo and task.
• Default final path: skills/<skill-name>/. If you staged the draft elsewhere, move the entire folder into skills/<skill-name>/ before declaring done.
• Keep SKILL.md lean: trigger boundary, workflow, decision rules, output contract, guardrails, and reference routing.
• Reuse existing references instead of duplicating them.
• Add new files only when clearly necessary and explicitly routed.
• Read references/authoring/description-engineering.md to craft the description field.
• During drafting, focus on the key constraints: SKILL.md under 500 lines, every reference file routed, description follows the formula. Run the full references/checklists/master-checklist.md review in Step 9.

8. Test the skill

• Trigger tests: Write 5+ should-trigger queries and 5+ should-NOT-trigger queries. For revisions, run each one by pasting it as a new message in Claude.ai (with only this skill enabled) or Claude Code, and record whether the skill loaded. For new skills, verify the description against your test queries manually; run live trigger tests after installation. See references/authoring/testing-methodology.md for the full testing guide.
• Functional test: Run at least one complete functional test of the primary workflow end-to-end.
• Self-check: Ask Claude "When would you use [skill-name]?" and verify the answer matches your intent.
• RED-GREEN-REFACTOR for discipline skills: If the skill enforces a rule an agent could rationalize away (TDD, verification, research-before-synthesis), run at least one pressure scenario without the skill first, capture the rationalizations verbatim, then re-test with the skill loaded. See references/authoring/tdd-for-skills.md for the full cycle.

Tip: For a NEW skill, install the draft to the active runtime's skill directory before testing triggers. Trigger tests fail silently if the skill isn’t loaded. For REVISIONS, test against the currently installed version. If the runtime forbids writing to the installed skill directory, run manual trigger review plus a functional workflow test, report the installation block explicitly, and do not claim live trigger coverage.

9. Finish with repo-fit and cleanup checks

• Confirm directory name, frontmatter name, and canonical labels stay aligned.
• Confirm every file in references/ is explicitly routed from SKILL.md.
• Confirm the result feels synthesized from evidence, not copied from the most detailed source.
• Confirm no < or > in frontmatter, no reserved names ("claude", "anthropic").
• If a shared repo issue blocks a clean result, report it instead of editing shared files by default.

Decision rules

• If outside research would materially change the answer, do not stay local.
• If the local workspace already contains the needed detail, route to that file instead of rewriting it.
• If two sources conflict, prefer the pattern that is clearer, leaner, and more compatible with the target repo.
• Prefer patterns that improve agent decisions. Reject bulk that adds machinery without improving the outcome.
• If the draft starts resembling one source too closely, stop and convert the source notes into inherit/avoid decisions before rewriting.
• If the skill grows because of examples, templates, or long checklists, move that detail to references or cut it.
• If remote research is skipped, say why it was safe to skip.
• If the skill enhances an MCP, verify tool names against the MCP server documentation before shipping.
• If the description is broad enough to fire on unrelated queries, add negative triggers.

Do this, not that

Output contract

Unless the user wants a different format, show work in this order:

1. skill type classification (after Step 2)
2. workspace scan summary (after Step 3)
3. source shortlist with selection rationale, or explicit skip reason (after Step 4)
4. markdown comparison table (after Step 5, required for the full research path)
5. success criteria (after Step 6)
6. selection and synthesis strategy (after Step 7)
7. generated or revised skill artifacts (after Step 7)
8. trigger test results (after Step 8)
9. final repo-fit and cleanup checks (after Step 9)

Reference routing

Load the smallest relevant set for the branch of work you are in.

Authoring

Patterns

Research

Quality and examples

Iteration and troubleshooting

Distribution

Guardrails and recovery

• Do not draft the final skill before reading the current workspace.
• Do not say research is complete unless the comparison table with inherit/avoid decisions has been produced for the non-trivial path.
• Do not hide the comparison table inside prose or skip it because sources "look similar."
• Do not copy a source skill wholesale, even as a temporary scaffold.
• Do not add files to references/ unless they are needed and routed.
• Do not ship without running at least basic trigger tests.
• Do not use < or > in any frontmatter field.
• Do not use "claude" or "anthropic" in the skill name.
• Do not reference bundled scripts with absolute paths. Script paths stay relative to the skill directory root.

Recovery moves:

• Local evidence is thin or contradictory: say so, then widen the evidence set before drafting.
• Research corpus is noisy: cut it down to the few sources you can actually compare well.
• Draft is bloating: move detail to existing references or delete it.
• Triggers are misfiring: read references/authoring/description-engineering.md and fix.
• MCP calls failing: read references/iteration/troubleshooting.md for diagnosis.
• Repo-fit conflict touches shared files: report the shared issue instead of editing outside the target skill by default.

Final checks

Before you finish, confirm all of the following:

• frontmatter name matches the directory name exactly
• name does not contain "claude" or "anthropic"
• description follows the formula: what + when + trigger phrases
• description is under 1024 characters with no < or >
• SKILL.md body is under 500 lines
• SKILL.md contains decisions and routing, not bulky reference content
• every reference file is explicitly routed
• routing verification: for f in $(find references -name '*.md' -type f); do grep -q "$(basename $f)" SKILL.md || echo "ORPHAN: $f"; done
• the final result is clearly synthesized from evidence and comparison
• no dead weight was added just because another skill had it
• at least 3 trigger tests passed (should-trigger and should-NOT-trigger)
• primary workflow completes without error in at least one test
• for full quality: ran references/checklists/master-checklist.md

Quick routing verification:

# Check for orphaned reference files (not mentioned in SKILL.md)
for f in $(find references -name '*.md' -type f); do grep -q "$(basename $f)" SKILL.md || echo "ORPHAN: $f"; done