Help us improve
Share bugs, ideas, or general feedback.
From skill-curator
Use when user asks to create, improve, or review production-grade SKILL.md files for Claude Code, Cursor, Codex, OpenCode, Kiro, and other agent tools. Focuses on trigger quality, structure, cross-tool compatibility, and reliable activation.
npx claudepluginhub agent-sh/skill-curator --plugin skill-curatorHow this skill is triggered — by the user, by Claude, or both
Slash command
/skill-curator:skill-curator [skill-purpose or --improve path/to/SKILL.md] [--category implementation|review|research|orchestration|analysis] [--minimal][skill-purpose or --improve path/to/SKILL.md] [--category implementation|review|research|orchestration|analysis] [--minimal]The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Expert guidance for writing high-quality, reliable `SKILL.md` files that work consistently across the entire agent ecosystem (Claude Code, Cursor, Codex, OpenCode, Kiro, Gemini CLI, and others).
Guides technical evaluation of code review feedback: read fully, restate for understanding, verify against codebase, respond with reasoning or pushback before implementing.
Share bugs, ideas, or general feedback.
Expert guidance for writing high-quality, reliable SKILL.md files that work consistently across the entire agent ecosystem (Claude Code, Cursor, Codex, OpenCode, Kiro, Gemini CLI, and others).
Use $ARGUMENTS as the requested skill purpose, improvement target, category, and scope flags:
$ARGUMENTS
--improve <path> means read the existing SKILL.md, critique it, then rewrite or patch it.--category implementation|review|research|orchestration|analysis sets the skill's primary workflow style.--minimal means keep the result compact and omit optional examples.Skills are the primary mechanism for giving agents specialized, on-demand knowledge without bloating every context window. A good skill:
agnix validation with zero errors/warningsEvery skill must start with a clean frontmatter block:
---
name: <kebab-case-name> # 3-8 words, lowercase, hyphens
description: <imperative trigger phrase> # 1-3 sentences, ≤ 512 chars
version: 1.0.1
argument-hint: "[optional args]" # Shown in /help and slash command
allowed-tools: Bash, Read, Edit, ... # Restrict when possible
---
The description is the primary activation signal. Models use it heavily for routing.
Requirements:
Good examples:
Bad examples:
Use this order for maximum model attention:
Skip unless: gates — mandatory for review/analysis skills)Every pattern or decision rule must contain a Skip unless: line that names a concrete, checkable condition (file name, function, identifier, command output, etc.).
Without this, the skill fires too broadly and generates noise.
Example:
### Missing error handling on async operations
The diff looks correct but is dangerous when:
- An async function is added or modified without try/catch or .catch()
Skip unless: the changed file contains `async ` or `await ` and the function name appears in a call site without error handling.
When a skill covers a broad domain, use the router pattern:
SKILL.md contains only frontmatter + a routing table.reference/specific-topic.md).See valkey-skills and agent-knowledge for production examples of this pattern.
Skills should work (at minimum) with:
.claude/skills/).cursor/skills/)Rules for broad compatibility:
allowed-tools in frontmatter to declare dependencies.agnix --target all or multiple targets.| Type | Target Lines | Hard Max | Notes |
|---|---|---|---|
| Simple skill | 40-80 | 150 | One clear responsibility |
| Standard skill | 80-150 | 250 | Most common |
| Router + references | 30-60 (main) + small refs | 400 total | Preferred for large domains |
| Heavy institutional memory | 150-300 | 500 | Only when justified by density |
Above 250 lines, the model starts ignoring later sections. Use routers instead of monolithic files.
Skip unless: gates for every decision rule.agnix on the resulting file (zero errors required).When asked to create or improve a skill, always return:
SKILL.md content in a clean code blockallowed-tools listSkip unless: requirement for pattern-based skills.(Generic skill-curator guidance. Source: agent-sh/skill-curator)