template

name

_template

description

TEMPLATE ONLY — not a real skill and never invoked. Copy this directory to .claude/skills/<your-skill-name>/, rename it, set `name` to match the new directory, and rewrite every section against docs/skill-generation-standard.md. Write the real description to the Portability contract: front-load what the skill DOES in the first ~90 characters, keep it ONE strict-YAML-valid line (single-quote it like this one and double internal apostrophes, 'like this'), parsed value under 1024 characters.

_template (reference skill)

This is a template, not a shipped skill. It lives at .claude/skills/_template/ and is deliberately ignored by scripts/validate-skills.py. Do not invoke it. To create a skill: copy this directory, rename it, and fill in every section below. Full rules: docs/skill-generation-standard.md.

Purpose

State, in one paragraph, what the real skill produces and the value it delivers. Example: "Produces a normalized spreadsheet and summary tab from a messy export."

Use When

• Use when: the user's request matches the concrete trigger the skill is built for.
• Use when: an adjacent-but-in-scope variation of that request appears.
• Do NOT use when: the request looks similar but belongs to another skill or a plain response — name that case explicitly so the model doesn't over-trigger.

Inputs to Inspect

• The repo files, docs, code, tests, and prior artifacts the real skill must read before acting. List them concretely so context precedes action.

Workflow

1. Gather and validate the inputs the skill needs (see Inputs to Inspect).
2. Perform the core transformation, reading references/ detail files on demand.
3. Produce the deliverable exactly as specified in Output Format.
4. Run the Validation Checklist before declaring done.

Output Format

Describe the exact deliverable: file path(s) and naming, report structure, or data schema. Be specific enough that two runs produce consistent shapes.

Validation Checklist

• Output matches the shape declared above.
• No Stop Conditions were silently bypassed.
• Any side-effecting step was gated behind explicit human confirmation. Under standard §5, the write an auto-invocable skill may make by default (Exception 1) is the approved doc/state append: create/append a non-executable documentation or project-state file, AFTER showing the exact path + content and receiving explicit, content-specific, single-use approval — path and content unchanged between approval and write. Autonomous source/test editing is NOT implied here; it exists only inside a separately classified and activated TALI route (Exception 2, §5), never by default.
• Any autonomous TALI validation command produced NO unapproved repository working-tree output (no cache/coverage/build/generated/ignored/temporary/report/ snapshot files outside the approved set), and any host-managed scratch stayed outside the repository and was removed before closeout.

Gotchas

• Keep SKILL.md under 500 lines; push detail into references/.
• Keep the frontmatter description trigger-oriented and under 1024 chars.
• Directory name and frontmatter name must stay identical.

Stop Conditions

• Stop and ask if required input is missing or the request is ambiguous.
• Stop and confirm before any irreversible or destructive action (delete, deploy, overwrite, spend).
• Before a documentation/project-state append under the standard-§5 approved-write exception (Exception 1): show the exact target path and exact content, wait for the explicit content-specific yes, and write only what was approved. A declined, ambiguous, or changed approval means NO write.
• Every side effect outside Exception 1 remains manual-only (disable-model-invocation: true) unless it is an ordinary approved source/test edit inside a separately classified and separately activated TALI route with every §5 precondition satisfied. Delete or rename; Git mutation; package installation; network, API, webhook, or MCP access; database write or seeding; deployment, provisioning, spending, or live-state mutation; protected-surface change; security-control change; governance change; architecture change; and every other §5 FORBIDDEN UNDER TALI class remain forbidden and must hand off to the owning policy. A generic approval does not convert any forbidden operation into a TALI operation. This template activates no route.

Supporting Files

• evals/evals.json — required for every real skill; the copy in this template directory (evals/evals.json) is the canonical starting point.
• evals/trigger-evals.json — add when the skill's trigger overlaps another skill.
• references/, assets/, scripts/ — add only when they reduce errors (progressive disclosure). Use "None" here if the skill is self-contained.