CommunityWriting & Editinggithub.com

AradhanaID/agent-skills

Write or rework a skill using the four-check method - trigger, structure, steering, pruning.

What is agent-skills?

agent-skills is a Claude Code agent skill that write or rework a skill using the four-check method - trigger, structure, steering, pruning.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/AradhanaID/agent-skills/tree/HEAD/skills/write-a-skill

Installed? Explore more Writing & Editing skills: steipete/notion, langchain-ai/langchain, bytedance/podcast-generation · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Write a Skill

One run authors or reworks one skill. The output is judged by behavior when the skill runs, not by how it reads: this run installs the user's process, not generic skill boilerplate.

Before writing, follow the context pointer in reference.md — frontmatter mechanics live there.

Steps

  1. Fix the trigger. Ask who must reach the skill and when. A skill that only ever runs on purpose is user-invoked (disable-model-invocation: true, description becomes a one-line human summary, <100 chars). A skill the agent must reach unprompted is model-invoked: its description is the always-loaded pointer, and its wording decides whether the fire happens. Done-when: the invocation choice names the load it is paying — context, or your memory.
  2. Inventory how the work actually runs. Recall the last real run of this work: what was sent back, corrected, or repeated. Those corrections are the skill's content; what the model does unaided is not. Done-when: the draft cites at least one real failure it prevents.
  3. Write steps, then pointers. Each step is one procedure with an explicit done-condition, ordered so the run ships something checkable first. Supporting material that only some passes need goes in reference.md behind a pointer naming when to read it. Done-when: every step is verifiable without asking the agent whether it succeeded.
  4. Pick the leading words. For each recurring idea, take the compact word already living in the model's pretraining and let it carry the meaning: the pointers, steps and done-conditions reuse those words. Rename jargon into them rather than explaining around them. Done-when: each leading word appears in more than one place and steers in each.
  5. Prune with the no-op test. Sentence by sentence: delete it, ask what broke. Nothing broke — the sentence goes. Whole sentences delete, never trimmed words; duplication collapses into a single source of truth rather than two fades. Done-when: you can no longer run the no-op test without passes; the file is shorter than the last version.

Reference

  • reference.md — frontmatter mechanics per invocation type, the pointer-writing rules, and the checklist battery for reviewing an existing skill.
  • The skill this run produces is iterated against real runs, not written once: when a run misbehaves, the read reference qualifies the failure before any rewrite.

Individual skills in this repo

This repo contains 3 individual skills — each has its own dedicated page.

Related Skills