AgentSkillsCN

new-skill

以交互式方式创建新的Claude Code技能。

SKILL.md
--- frontmatter
name: new-skill
description: Interactively create a new Claude Code skill
disable-model-invocation: true
argument-hint: "[skill-name]"

Walk the user through creating a new Claude Code skill step by step.

Step 1 — Name and scope

If $ARGUMENTS is provided, use it as the skill name. Otherwise ask the user for a name (lowercase, hyphens allowed).

Ask whether this should be a personal skill (~/.claude/skills/<name>/) or a project skill (.claude/skills/<name>/).

Step 2 — Gather requirements

Ask the user these questions (use AskUserQuestion):

  1. •What should the skill do? — get a plain-language description of the desired behavior.
  2. •Who triggers it?
    • •User only (disable-model-invocation: true) — good for actions with side effects
    • •Claude only (user-invocable: false) — background knowledge Claude applies automatically
    • •Both (default) — Claude can invoke it when the description matches
  3. •Does it need arguments? — if yes, ask for an argument-hint (e.g. [filename], [issue-number]).
  4. •Should it run in isolation? — if the skill does heavy read-only analysis, suggest context: fork with an appropriate agent type.
  5. •Should tool access be restricted? — if the skill should be read-only or limited, ask which tools to allow.

Step 3 — Write the SKILL.md

Using the answers, generate the SKILL.md file with:

  • •Correct YAML frontmatter (name, description, and any optional fields gathered above).
  • •Clear, concise markdown instructions (under 500 lines).
  • •Use $ARGUMENTS / $0 / $1 where the user wants input substitution.

Create the directory and write the file.

Step 4 — Trigger analysis

Before finalizing, carefully review the skill for triggering issues. Consider and briefly report to the user:

  • •False positives — Is the description broad enough that Claude might invoke the skill when the user didn't intend it? If so, tighten the wording or add disable-model-invocation: true.
  • •False negatives — Would a user who wants this skill use phrasing that doesn't match the description? If so, broaden the description or add synonyms.
  • •Collisions — Could the description overlap with other common skills (e.g. a "review" skill clashing with PR review)? If so, make it more specific.

Suggest concrete revisions if any issues are found. Apply them with user approval.

Step 5 — Token efficiency review

Review the skill's instructions and identify steps where Claude would spend many tokens doing work that a shell script or CLI command could handle instead. Common opportunities:

  • •Listing/searching files → find, grep, glob in a script
  • •Parsing structured data (JSON, YAML) → jq, yq
  • •Gathering git context → a small shell snippet in the instructions
  • •Repetitive formatting → a template file or sed/awk
  • •Heavy analysis → claude -p as a subprocess with a focused prompt

Present any suggestions to the user. If they agree, update the skill instructions to delegate that work to commands.

Step 6 — Confirm

Show the user the final file contents and the slash command they can use to invoke it. Mention they can test it immediately with /<name>.