AGENT SKILL

build-skill

Create, redesign or merge a skill, grounded in facts, then hand over the test. Started by you.

License 许可
MIT
最近更新
2026-10-26
Stars
★ 0
Forks
0

01 前言 Front Matter

name
build-skill
description
Create, redesign or merge a skill, grounded in facts, then hand over the test. Started by you.
allowed-tools
Bash(agentic-maker *) Bash(${CLAUDE_PLUGIN_ROOT}/bin/agentic-maker *) Bash(claude plugin validate *) Read Grep Glob Write Edit
argument-hint
<what the skill should do | path of a skill to redesign>
disable-model-invocation
true

02 资源文件 13

03 SKILL.md 正文

Build a skill

In Codex: $ARGUMENTS is what the user wrote after the skill name. ${CLAUDE_PLUGIN_ROOT} and ${CLAUDE_SKILL_DIR} mean the plugin folder and this skill's folder (the plugin folder is two levels above this file); agentic-maker is on PATH or is <plugin folder>/bin/agentic-maker. A line that starts with an exclamation mark and a backtick is a command: run it yourself first.

Request: $ARGUMENTS

Pages are in ${CLAUDE_SKILL_DIR}/references/ (listed at the end); the two skill pages in the knowledge base are ${CLAUDE_PLUGIN_ROOT}/skills/agent-interface-design/references/skill/. Open a page when the step names it, not before. This runs in the main conversation, so ask the user directly, and only what changes the result.

Pick the mode

The request is Do
a new skill steps 1-7
an existing skill to redesign (a path, relative to the current working directory, not to the plugin) ls that path in the current directory and read it first. If the user has derail findings for it, read them; they say how it fails today. Then steps 3-7
several skills or sources to merge step 2 builds a comparison table (what to inherit, what to drop, why) before any text is written. Every command, flag and path comes verbatim from a source; one found in no source goes in the report as a suggested addition, not in the skill
"turn this conversation into a skill" list what was done as a nested outline, five levels at most: milestone, phase, step, error met and its correction, invariant learned. Corrections and invariants become the skill's rules; repeated commands become scripts. Then steps 2-7

Steps

  1. Scope. Who invokes it (the model on a matching request, or only the user), in which harnesses it must run, and one sentence for what changes for the agent when it is loaded. naming-and-boundaries.md if one skill or two.
  2. Ground it. For any non-trivial domain, gather verified facts first: the official docs, the real command output, the API as it is today. If the research plugin is installed use /research:ask (one question) or /research:deep (five or more sources); otherwise use whatever web and file tools you have. A skill drafted from memory carries the model's guesses into every future run; if you can verify nothing, say so and stop. Look at two or three existing skills for the same job and note what each does well.
  3. Lay it out before writing. What stays in SKILL.md (under 500 lines), what goes to references/, what becomes a script: references-and-scripts.md. Pick the body's shape: workflow-patterns.md. Show the layout to the user when the request was open-ended.
  4. Write it.
  5. Check what a script can check. Run each check as its own Bash call, with no &&, ; or heredoc: claude plugin validate <skill dir> (frontmatter and structure), wc -l <skill dir>/SKILL.md, every bundled script on real input. A denied call skips only that check: redo it with Read, Grep or Glob (frontmatter fields, line count, each link and path) and name the denied command in the report. Then open checklist.md and build a table with one row per line, marked checked or unchecked.
  6. Hand over the test. Only the user can start /agentic-maker:derail; you cannot, so do not try. Write three scenarios into .agentic-maker/scenarios/<name>.md (format in ${CLAUDE_PLUGIN_ROOT}/skills/derail/references/scenarios.md): a normal request in the user's words that never names the skill, an ambiguous one, and one with a missing input; each with receipts that check the disk, not the agent's words. Preview one with agentic-maker run <name> --dry-run. End by giving the user the exact command, /agentic-maker:derail <skill dir>. Until they run it the skill is untested, and your report says so. Procedure and what a pass looks like: test-with-fresh-agents.md. A skill only the user can start is tested through a slash-command prompt; the derail dry-run says so.
  7. Ship. The final message is the checklist table from step 5: one row per line of checklist.md, marked checked (with the command and its output, or the file and line you read) or unchecked. A message without that table is not a ship.

Do this, not that

Do Not
Verify every API, flag and path against the real thing Write them from memory
Fix the sentence that misled the agent Add a "never do X" rule next to it
Keep SKILL.md as the routing and the rules; move bulk to references Paste a 1,500-line guide into one file
Write scripts in Python or Node and run them before you ship Ship a script you never ran
Report what you could not verify Present a guess as a fact

Pages

Open When
frontmatter.md writing or fixing frontmatter
naming-and-boundaries.md naming it, splitting or merging
references-and-scripts.md deciding what goes where
workflow-patterns.md choosing the structure of the body
instruction-strength.md a step is too loose or too rigid
wrapping-mcp-tools.md the skill calls MCP tools
test-with-fresh-agents.md step 6
checklist.md before shipping
troubleshooting.md it does not load, trigger or behave
anti-patterns-content.md, anti-patterns-process.md reviewing a draft or your own process
worked-examples.md you want a structure to copy

04 同仓库更多 Skills