AGENT SKILL

skill-factory

Create a new re_gent skill from a plain-language description, writing a valid SKILL.md. Use when the user wants a new skill, or to build one that style-factory proposed.

License 许可
Apache-2.0
所属仓库
bonez-io/re_gent
最近更新
2026-11-11
Stars
★ 5
Forks
0

01 前言 Front Matter

description
Create a new re_gent skill from a plain-language description, writing a valid SKILL.md. Use when the user wants a new skill, or to build one that style-factory proposed.
allowed-tools
Bash(rgt log *), Bash(rgt sessions *), Glob, Read, Write
argument-hint
<description of what the skill should do>

03 SKILL.md 正文

Turn a description into a working skill file.

A skill is a prompt plus a tool grant. The intelligence is the agent's, the data is re_gent's, and the skill is the wiring between them — so writing one is mostly deciding what to ask for and what it is allowed to run.

1. Settle the shape

Before writing, be able to answer in one sentence each:

  • Trigger — when should an agent reach for this? This becomes description, and it is the single most important line in the file: it is what the agent matches against. Name the situation, not the mechanism.
  • Input — a path, a step hash, a symptom, nothing? This becomes argument-hint.
  • Data — which rgt commands supply the facts.
  • Output — what the user should be holding at the end.

If the request is vague, ask one question. A skill built on a guess triggers at the wrong moment and is worse than no skill.

2. Check it is not already there

Use Glob to list .claude/skills/*/SKILL.md (or the active host's matching skills directory), then read any neighbour that looks close.

Extending an existing skill beats adding a near-duplicate — two skills with overlapping descriptions make the agent pick badly.

3. Write it

.claude/skills/<kebab-case-name>/SKILL.md:

---
description: <what it does + when to use it, one sentence>
allowed-tools: Bash(rgt <command> *)
argument-hint: "<arg>"
---

<Instructions to the agent.>

Rules that matter:

  • Grant the narrowest tools that work. Bash(rgt log *) — never bare Bash. A skill that only reads should never be able to write.
  • Write instructions, not prose. The reader is an agent about to act.
  • Say what to do with the data, not just how to fetch it. The gathering is the easy half; the analysis is the skill.
  • Prefer real commands you have verified exist. Check with rgt <cmd> --help before referencing a flag. A skill that cites a flag that does not exist fails at the moment it is trusted.
  • Name sibling skills the agent should chain to, so skills compose.

4. Verify before claiming it works

rgt <each command in allowed-tools> --help

Confirm every command and flag exists. Then state plainly what you could not test — a skill's real behaviour only shows when an agent runs it against real history.

Report

Path written, the description line verbatim (so the user can judge when it will trigger), the commands it may run, and one example invocation.

04 同仓库更多 Skills