generate-skill

Generate Skill

Overview

Produces one ready-to-ship SKILL.md (plus optional references/, scripts/, assets/) on the conventions from the agentskills.io skill-creation pages, Anthropic's skill-creator, the platform prompting guides, and empirical activation research. The Phase 4 calibration list is tuned for the Claude 5 generation; re-check it against the current model's prompting guide when the generation changes.

One skill, one job. This skill's job is the SKILL.md and its frontmatter — not docs, releases, or auxiliary files.

Two principles drive every phase below:

• The description is the product. It is the only thing Claude reads to decide whether to invoke the skill. Tune it like a prompt.
• Add what the agent lacks, omit what it knows. For every instruction ask: would the agent get this wrong without it? If no, cut it. Encode only non-inferable, procedural, skill-specific knowledge.

Navigation

Workflow

Use AskUserQuestion, one question at a time — never bulk-dump the questionnaire. After each phase show the artifact and confirm before moving on.

Phase 1 — Discovery

1. Purpose. One sentence: what job does this skill do?
2. Type, which drives the body template: methodology (multi-step workflow — code-review, track-session) · technical (an API, format, or tool — docx, semantic-release) · auditing (grades an artifact — rate-skill, security-review) · reference (schemas, conventions, lookup tables — bigquery) · automation (a script or external command — screenshot-local).
3. Trigger phrases. 5+ verbatim phrases a real user would say. Push for naturalese: "create a skill", not "scaffold skill artifact".
4. Negative scope. Which near-neighbors should this not steal from? Drives the Do NOT use… clause and the should-not-trigger half of the Phase 8 eval set.
5. Enforcement level. Suggestion, guided (checklists), or strict (gated phases). Defaults: methodology and auditing = guided, technical/reference/automation = suggestion.

Phase 2 — Description drafting

Official sizing is "a few sentences to a short paragraph," hard cap 1024 chars. Shape:

<Third-person noun phrase>. Use whenever the user <wants/asks to> <trigger 1>, "<trigger 2>", or <trigger 3> — even if they don't explicitly mention <domain term>. Do NOT use this skill for <near-neighbor> — see <other-skill>.

Show the draft, measure it (python3 scripts/measure.py <path/to/SKILL.md> reports description chars, body lines and estimated body tokens against their caps), iterate.

• ✅ Distinctive noun in the first 50 chars
• ✅ Imperative third person addressed to the agent — the skill, not "I"
• ✅ At least one literal quoted user phrase
• ✅ Coverage clause for indirect asks ("even if they don't explicitly mention X") — officially recommended
• ✅ Do NOT use this skill for … when a near-neighbor exists
• ❌ First person ("I help you create skills") — degrades activation (Seleznov n=650)
• ❌ Vague ("A skill for creating skills"), or a bullet-list description — no surveyed top skill uses that format
• ❌ Intensity escalation ("CRITICAL: you MUST invoke this") — the model overtriggers on aggressive language, and the official fix is normal register. Raise recall with more literal triggers and a coverage clause, never with volume.

Phase 3 — Frontmatter

Emit these fields, in this order, and nothing else unless the skill needs it:

---
name: <kebab-case, ≤64 chars, no "anthropic"/"claude" reserved words>
description: <string from Phase 2>

effort defaults by type: automation and reference low, technical medium, auditing and methodology high, multi-source synthesis and agentic coding xhigh. A skill whose deliverable is a long file stays at high — at xhigh/max the model drafts it twice, once in reasoning and once as output. Omit the key to inherit; rationale and the live-verification note are in FRONTMATTER.md.

Other optional fields, frontmatter anti-patterns and the three-tier portability rules: references/FRONTMATTER.md.

Phase 4 — Body content

Full templates in references/PATTERNS.md — load only when generating the body. This section list is the shared spec graded by rate-skill Category 4; keep the two skills in sync.

Content rules, every type:

• Always include ## Gotchas — canonical: "the highest-value content in many skills is a list of gotchas." Keep them in SKILL.md, never a reference file, because the agent must read them before hitting the situation. Every correction you have to make to the deployed skill's agent becomes a new gotcha.
• Five more official instruction patterns beyond gotchas: output-format templates (agents pattern-match concrete structures better than prose descriptions of a format), checklists for multi-step workflows, validation loops (do → validate → fix → repeat), plan-validate-execute for batch or destructive operations, bundled scripts where the agent would otherwise reinvent the same logic each run. Use the ones that fit, not all of them.
• Provide defaults, not menus — pick one approach, mention alternatives briefly.
• Favor procedures over declarations — teach how to approach the class of problem, not what to output for one instance.
• Pair every "do not X" with a positive directive. Official: "Tell Claude what to do instead of what not to do."

Agentic calibration — leave these out. The model already does them, so restating the instruction compounds the behavior and burns tokens with no quality gain.

• No generic self-re-check steps — cut "double-check your answer", "re-verify before responding", "add a final verification step for any non-trivial task". The fix is deletion, not rewording. Three things stay, and the last two are recommended for long-running work: checks against external state (run the test suite, confirm the file parses, validate against the schema); a writer-verifier pattern where a fresh-context agent judges another agent's output; and an evidence audit before a progress report ("only report work you can point to a tool result for"). The defect is an agent re-reading work it produced itself with nothing external to check it against.
• Never instruct the agent to echo its reasoning ("show your thinking", "explain your reasoning in the response") — can trigger the reasoning_extraction refusal and a model fallback. Ask for conclusions and evidence instead.
• Never instruct the agent not to think or not to reason — increases leakage of internal XML tags into visible output. Say nothing about reasoning at all; shape the output instead ("one chat line, the file is the deliverable").
• No narration suppressors ("hold all findings for the final response", "don't narrate", "no interim updates") and no anti-formatting rules ("never use bullets", "no headers", "no bold"). Both were written against models that over-narrated and over-formatted; the model does the opposite, so the lines produce silence and flat prose. Say when a specific update or format is wanted instead ("report scope before dispatch"; "use a table when comparing three or more items"). Shaping the final message — one chat line, the file is the deliverable — is fine.
• Scope delegation when the skill uses subagents: name the scenarios that warrant one and the cap, dispatch independent agents in one message, and keep working while they run instead of blocking on each. Open-ended delegation multiplies cost; blocking delegation multiplies wall-clock.
• Bound the deliverable when the skill writes a file — the output template needs a line like "match length to the task, no filler sections, redundant summaries or boilerplate." Written deliverables run long by default.
• State where a narrow skill stops — models expand scope on their own, adding steps nobody requested. When the skill produces code, add the code-side line too: "Report a pre-existing bug or behavior the task doesn't mention as a follow-up rather than fixing it here. Commit tests only where the task asks or the repo already keeps tests for this kind of change, sized like the neighboring test files; scratch checks are not test files." Unrequested additions and committed test code drop substantially with this line, with no change in task success.
• Don't restate what the harness injects — the autonomy block, the delivering-work scope block, the progress-update line, the overplanning nudge, the hidden-tool-output note, the batch-tool-calls nudge, the readability rules are already in every Claude Code session; a second wording makes the model reconcile the two.

Phase 5 — Examples

Produce 3–5 covering distinct situations, not one situation restated — vary them enough that the agent doesn't pattern-match an unintended regularity. At least one is a ❌ counter-example. Show ✅ first; if room, end on ✅ (recency bias). Each gets a one-line verdict:

Use ✅/❌ for the quality labels — the <Good>/<Bad> XML form appears in zero of the 8 surveyed top community skills. This is about the labels, not XML: <example>/<examples> as structural delimiters are officially recommended and fine. Once the body passes 300 lines these move to references/EXAMPLES.md behind a pointer naming what's in it; that counts in full for grading.

Phase 6 — Gotchas

Failure modes specific to this skill, not generic skill-authoring advice. Each entry: one-line symptom, one-line cause, one-line fix.

Phase 7 — Progressive disclosure check

Measure with scripts/measure.py. Extract once SKILL.md passes 300 lines; the joint 500-line / ~5,000-token cap is the hard ceiling either way.

1. List candidate sections to extract, largest first.
2. Propose references/<TOPIC>.md files. Plural is the spec-documented name, but nothing validates directory names — never rename an existing singular reference/ dir for style.
3. Keep extracted files exactly one level deep; deeper nesting gets silently truncated by preview reads.
4. Give each a one-line pointer in SKILL.md: when to load it, what's in it. Exception: gotchas stay.

Phase 8 — Eval set (required before finalize)

Build the set the official description-optimization loop consumes:

1. ~20 queries — 8–10 should-trigger (Phase 1 phrases plus paraphrases and indirect asks) and 8–10 should-not-trigger (Phase 1 near-neighbors plus unrelated work).
2. Split 60/40 train/validation, proportional mix in both halves. Shuffle once, then keep the split fixed across iterations.
3. Save as references/EVAL.md with the split marked and the protocol stated: each query run in a fresh session ~3 times; trigger rate = fraction of runs that invoked the skill; should-trigger passes above 0.5, should-not below.
4. Iterate on train failures only — never peek at validation to choose wording. Check validation after and select the best validation score, since later iterations overfit and an earlier draft can win. "Five iterations is usually enough."
5. Have the user spot-check 3–5 queries in a fresh Claude session before finalizing.

Caveat (official): agents consult skills for tasks beyond what they handle alone, so a trivially simple query may not trigger even a perfectly described skill. Don't churn the description over those.

To measure whether the body improves output rather than just triggering, run the official output-quality loop: 2–3 test cases with and without the skill, graded on evidence-required assertions written after the first run, dropping any assertion that passes in both configurations. Worth it for methodology and auditing skills, overkill for small reference skills.

Output layout

Examples

✅/❌ pairs for methodology and technical descriptions, a counter-example with its four defects reversed, and body steps that check external state vs. re-check the agent's own work: references/EXAMPLES.md.

Gotchas

• Symptom: New skill never auto-invokes. Cause: Vague prose, no specific triggers. Fix: Rewrite in the "Use whenever the user wants to…" form with 3+ quoted trigger phrases.
• Symptom: Works in isolation, breaks once the user has >20 skills installed. Cause: Distinctive trigger sits past char 50 and the listing budget truncated it. Fix: Move the distinctive noun to the start of description.
• Symptom: Fires on exact phrases but misses indirect asks ("clean up this data file" for a CSV skill). Cause: No coverage clause. Fix: Add "even if they don't explicitly mention X".
• Symptom: Description scores well on train, regresses on validation. Cause: Overfitting — wording tuned against the same queries each round. Fix: Select by validation score, keep the split fixed; an earlier iteration may win.
• Symptom: "Unexpected key" warning on load. Cause: Top-level version, author, or tags. Fix: Move under metadata. argument-hint and hooks are valid top-level Claude Code keys and stay put — only Anthropic's packaging validator rejects them.
• Symptom: Two skills both fire on one prompt. Cause: Overlapping triggers, no negative scope. Fix: Add Do NOT use this skill for X — see Y to whichever is the wrong fit.
• Symptom: SKILL.md is 700 lines and the agent quotes the wrong section. Cause: Single-file overflow — Claude reads the head and misses the tail. Fix: Extract to references/, one level deep, with explicit "load when…" pointers.

Integration

• rate-skill — run after generating to grade the result. generate-skill produces, rate-skill audits; the Phase 4 table and rate-skill's Category 4 table are one shared spec.