Skill Curator

Skill Curator

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).

Parse Arguments

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.

Why Skills Matter

Skills are the primary mechanism for giving agents specialized, on-demand knowledge without bloating every context window. A good skill:

• Activates reliably when relevant (strong trigger phrases)
• Loads only the necessary content (router pattern or tight scope)
• Survives model changes and tool updates
• Passes agnix validation with zero errors/warnings
• Provides clear, actionable guidance with "Skip unless" gates

Frontmatter Standards

Every 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
---

Description (Trigger Phrase) Rules — Most Important Field

The description is the primary activation signal. Models use it heavily for routing.

Requirements:

• Start with an imperative verb phrase: "Use when...", "Write when...", "Review when...", "Analyze when..."
• Be specific — mention file paths, function names, command names, or concrete situations
• Use vocabulary that will actually appear in user prompts or PR/issue text
• Include negative signals ("Does NOT apply to...") when scope could bleed
• Keep under 512 characters

Good examples:

• "Use when adding or modifying background job queues, workers, or schedulers in Node.js or TypeScript projects."
• "Review when a PR touches cluster replication, failover logic, or handshake code in the C server."
• "Use when the user wants to create or improve a production-grade SKILL.md file for any agent tool."

Bad examples:

• "Helps with background jobs." (too vague, no trigger)
• "I am an expert in queues." (first person, not imperative)

Body Structure (Recommended Order)

Use this order for maximum model attention:

1. Purpose (short)
2. Core Principles or When to Use
3. Frontmatter Requirements (if relevant)
4. Template / Structure
5. Patterns (with Skip unless: gates — mandatory for review/analysis skills)
6. Do NOT / Anti-patterns
7. Workflow (step-by-step for the skill's task)
8. Examples
9. Constraints / Token Budget

The "Skip unless" Rule (Critical for Reliability)

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.

Router Pattern (For Complex Skills)

When a skill covers a broad domain, use the router pattern:

• The main SKILL.md contains only frontmatter + a routing table.
• It loads small, focused reference files on demand (reference/specific-topic.md).
• This keeps initial context tiny while still providing deep knowledge.

See valkey-skills and agent-knowledge for production examples of this pattern.

Cross-Tool Compatibility

Skills should work (at minimum) with:

• Claude Code (.claude/skills/)
• Cursor (.cursor/skills/)
• Codex CLI
• OpenCode
• Kiro
• Gemini CLI (where supported)

Rules for broad compatibility:

• Avoid Claude-specific tool names in the skill body unless gated.
• Prefer standard Bash + file tools when possible.
• Use allowed-tools in frontmatter to declare dependencies.
• Test with agnix --target all or multiple targets.

Length Budgets

Above 250 lines, the model starts ignoring later sections. Use routers instead of monolithic files.

Common Failure Modes (and Fixes)

• Vague description → Never activates. Fix: Make it imperative and specific with real tokens from the domain.
• No "Skip unless" → Fires on everything. Fix: Add concrete gate conditions for every pattern.
• Buried rules → Model misses them. Fix: Put critical constraints (confidence ladder, always-on checks, Do NOT) near the top.
• First-person voice → Breaks selection in some tools. Fix: Use third-person / imperative throughout.
• Tool-specific assumptions → Breaks on other platforms. Fix: Declare tools in frontmatter and keep body generic where possible.
• Missing examples → Agent doesn't know the expected output shape. Fix: Include at least one realistic before/after or full trajectory.

Workflow When Using This Skill

1. Clarify the skill's purpose and target tools.
2. Decide on scope (single-purpose vs router).
3. Draft frontmatter with an excellent description.
4. Choose structure and write the body following the template above.
5. Add Skip unless: gates for every decision rule.
6. Run agnix on the resulting file (zero errors required).
7. Test activation with realistic prompts in the target tools.
8. Add to the appropriate skill registry or plugin.

Output Requirements

When asked to create or improve a skill, always return:

• The complete SKILL.md content in a clean code block
• A short critique of the previous version (if improving)
• Token estimate for the frontmatter + body
• Recommended allowed-tools list
• Suggested test prompts that should trigger it

Constraints

• Never write vague or marketing-style descriptions.
• Never omit the Skip unless: requirement for pattern-based skills.
• Never assume a specific tool's internal implementation details unless the skill is tool-specific.
• Always optimize for reliable, low-noise activation over cleverness.

(Generic skill-curator guidance. Source: agent-sh/skill-curator)