Agent Skills

Agent Skills Standard Reference

This skill documents the Agent Skills open format — a standardized way to give AI agents new capabilities and expertise. Follow this workflow when creating or editing skills in this repository.

Authoritative source: agentskills.io/specification. The bundled specification is a working snapshot; check the authoritative source when currentness matters.

Directory Structure

A skill is a directory containing, at minimum, a SKILL.md file:

skill-name/
├── SKILL.md          # Required: metadata + instructions
├── evals/            # Required for new skills in this repository
├── scripts/          # Optional: executable code
├── references/       # Optional: documentation
├── assets/           # Optional: templates, resources
└── ...               # Any additional files or directories

Required Workflow

Create or edit a skill

1. Read the specification before changing SKILL.md metadata or directory structure.
2. Ground instructions in real domain knowledge, project artifacts, and observed failure modes. Read best practices when designing or materially revising instructions.
3. Keep the skill a coherent, triggerable unit. Put only essential instructions in SKILL.md; put conditional detail in focused reference files and state exactly when to read each one.
4. Use a precise description that says what the skill does, when it applies, and when it does not apply. For skills with meaningful overlap, name the nearest alternative or prerequisite in a ## When not to use section. Test the boundary with at least three should-trigger prompts and two should-not-trigger near-misses; keep these harness-specific trigger checks separate from portable output-quality evals. Read optimizing descriptions for trigger design.
5. For every new skill, create evals/evals.json with at least five representative output-quality cases. Each case needs a realistic prompt, an expected outcome, and observable assertions. Include edge cases that exercise risky or ambiguous behavior. Read evaluating skills for the eval format and iteration workflow.
6. When bundling executable code, read using scripts. Document prerequisites and non-interactive invocation in the skill.
7. Before handoff, run the validation checks in this skill and correct every finding.

Review a skill

1. Validate the required frontmatter, field constraints, parent-directory/name match, and YAML syntax against the specification.
2. Check that the description has both positive and negative trigger boundaries, the workflow is actionable, and resource references are conditional and reachable.
3. Check that any script has documented dependencies, safe non-interactive inputs, clear errors, and structured output where useful.
4. For this repository, also verify its required human-facing README.md: title, Why Install This Skill, What You Get, Quick Start (unless genuinely reference-only), Triggers, and Requirements. Keep it human-facing, concise, and free of agent-only instructions.
5. Verify every new skill has evals/evals.json with at least five output-quality cases, and check that their expected outcomes and assertions test meaningful behavior. For an existing skill without evals, flag the gap but do not block the review solely for that legacy absence. Read evaluating skills before declaring the work complete.

Client implementation work

When implementing skill discovery, activation, or context management in an agent product, read client implementation guidance. Do not apply client conventions such as search paths as universal format requirements.

SKILL.md Format

The SKILL.md file must contain YAML frontmatter followed by Markdown body content.

Frontmatter Fields

name field rules

• 1–64 characters
• Only lowercase unicode alphanumeric (a-z, 0-9) and hyphens (-)
• Must not start or end with a hyphen
• Must not contain consecutive hyphens (--)
• Must match the parent directory name

description field rules

• 1–1024 characters
• Should describe both what the skill does and when to use it
• Should include specific keywords that help agents identify relevant tasks

compatibility field rules

• If present, it must be 1–500 characters
• Include it only for concrete environment requirements, such as a required product, system package, network access, or runtime

metadata field rules

• Must be a map of string keys to string values
• Use reasonably unique keys to avoid collisions with other clients or tools

Body Content

The Markdown body has no format restrictions beyond being helpful to the agent. Recommended sections:

• Step-by-step instructions — the procedure the agent should follow
• Examples of inputs and outputs — what data looks like going in and coming out
• Common edge cases — situations the agent might not handle correctly without guidance
• Gotchas — environment-specific facts that defy reasonable assumptions

Keep SKILL.md under 500 lines and 5000 tokens. Move detailed reference material to separate files in references/.

Progressive Disclosure

Agents load skills in three stages:

1. Metadata (~100 tokens): name and description loaded at startup for all skills
2. Instructions (< 5000 tokens recommended): Full SKILL.md loaded when activated
3. Resources (as needed): Files in scripts/, references/, assets/ loaded on demand

Supporting Directories

scripts/

Executable code agents can run. Scripts should:

• Be self-contained or clearly document dependencies
• Include helpful error messages
• Handle edge cases gracefully
• Use relative paths from the skill root (e.g., scripts/extract.py)
• Be explicit about intent: say "run this script" or "read this as reference" — never leave the agent to guess whether a bundled file is executable or illustrative

references/

Additional documentation loaded on demand. Keep individual files focused — agents load these when instructed, so smaller files save context.

• Every file under references/ must be at or under 60,000 characters. When a reference grows past the cap, split it into focused files (e.g., references/<topic>-a.md, references/<topic>-b.md) and update SKILL.md so each new file is reachable and the load-on-demand instructions name the right file.