Writing Skills

Steps

1. Clarify the trigger condition

Before writing anything, establish: in what situation should this skill activate? The answer must be a specific, observable event (something the user says or does), not a description of what the skill will accomplish. This becomes the when_to_use field, not description.

If the trigger is vague ("when it seems useful"), ask for a concrete example before proceeding.

2. Determine where the skill lives

| Scope | Location | |-------|----------| | Global (applies across all projects) | ~/.claude/skills//SKILL.md | | Project-specific | .claude/skills//SKILL.md in the project root |

Ask if not obvious from context.

3. Pick the frontmatter fields

Every skill uses name, description, and when_to_use. Go through the rest of the field list and include only what genuinely applies; omit the rest.

| Field | Include when | |-------|-------------| | description | Always. One sentence on what the skill produces or does; outcome-focused, not a walkthrough of the steps. | | when_to_use | Always. Trigger phrases, example user requests, and the anti-shortcut suffix: Follow all steps in order; do not shortcut based on this description. | | argument-hint | The skill accepts an argument (URL, file path, vault name). Format as "[what it accepts]". | | model: sonnet (or haiku) | The skill is a mechanical workflow that does not need the session model's full capability; pinning a cheaper model saves usage. Omit for high-judgment skills (review, interview, design) so they inherit the session model. Never pin a model to force quality upward: a pin caps the skill when the session runs a stronger tier (e.g. Fable). | | effort: high | The skill requires high-judgment reasoning and benefits from deeper thinking. | | allowed-tools | The skill runs repetitive commands that would otherwise prompt (e.g. Bash(git *) Bash(gh *) Read). Additive, not restrictive. | | paths | The skill naturally scopes to specific files expressible as a glob (e.g. **/*.config.json). Auto-triggers the skill when editing those files. | | disable-model-invocation: true | The skill has side effects that should require deliberate user invocation (posting to a live site, deploying, deleting). | | user-invocable: false | The skill is pure background knowledge; the user never types /skill-name. | | license | The skill is open-source or shareable. Common values: MIT, Apache-2.0. | | metadata | Custom key-value pairs for distribution: author, version, mcp-server, category, tags, documentation, support. Useful when publishing publicly. |

4. Draft the skill

Use this structure. Omit any field that does not apply.

---
name: 
description: ""
when_to_use: ""
argument-hint: "[...]"
model: sonnet
effort: high
allowed-tools: Bash(git *) Read
paths: "**/*.ext"
disable-model-invocation: true
---

## Steps

[The actual workflow. No summary at the top. Jump straight into the first step.]

Optional companion folders. A skill is a folder, not just a single file. When SKILL.md grows large or needs supporting assets, split content out:

/
├── SKILL.md            # required
├── scripts/            # optional — executable code (Python, Bash, etc.)
├── references/         # optional — detailed docs, loaded only when needed
└── assets/             # optional — templates, fonts, icons

Reference these from SKILL.md by path (e.g. See references/api-patterns.md for rate limiting guidance). Do not add a README.md inside the skill folder — all human-facing documentation belongs in SKILL.md or references/. A repo-level README is fine when distributing via GitHub; it just must not live inside the skill directory itself.

Rules for description:

• One sentence, outcome-focused.
• Good: "Reviews a PR and reports issues grouped by severity."
• Bad: "Gets the diff via gh, runs composer lint, writes numbered comments." (too implementation-specific; invites shortcut-following.)
• Never stuff trigger phrases into description; they belong in when_to_use.

Rules for when_to_use:

• Start with "Use when the user..." and list concrete triggers.
• Include the anti-shortcut suffix: Follow all steps in order; do not shortcut based on this description.
• description and when_to_use are concatenated in the skill listing and share a 1,536-character cap. Keep both tight.
• Add negative triggers when over-triggering is a real risk: Do NOT use for [adjacent case] (use [other-skill] instead). This is more effective than tightening the positive trigger phrasing alone.

Rules for the body:

• Do not open with a paragraph summarising what the skill does. Jump directly to Step 1.
• Steps should be imperative and specific. "Search for X before asking" is better than "You may want to consider searching for X."
• If parts of the workflow must be followed exactly (no adaptation), mark them explicitly: This step is not optional.
• For sections that must not be skipped, prefer a dedicated ## Important or ## Critical heading over inline bold. Headers are harder to skim past than emphasis inside a paragraph.

5. Pressure-test before finalising

Run two mental tests.

> If an agent read only description + when_to_use, what would it do?

The right answer is "decide whether this skill applies to the current task," not "attempt the task." If the metadata reads like a walkthrough, tighten the description.

> What would a lazy agent do if this skill did not exist?

Write down that failure mode. Check that the skill's steps directly prevent it. If not, add a step or a "What to avoid" section.

6. Present for approval

Show the full draft. Do not write the file until the user confirms.

7. Write the file

Once approved, write the SKILL.md to the chosen location:

• Global skills: ~/.claude/skills//SKILL.md.
• Project skills: /.claude/skills//SKILL.md.

The folder under skills/ must be a top-level directory, not nested. Use the frontmatter name: value as the directory name (e.g. skills/obsidian-charts/, not skills/obsidian/charts/), because Claude Code only scans the top level.

---

What to avoid

• Mixing what-it-does and when-to-invoke in description. description summarises the outcome; when_to_use carries the triggers.
• Writing a walkthrough in description. If it reads like step-by-step, compress to the outcome. Implementation details invite shortcut-following.
• Opening the body with an explanation of the skill. The first line of the body should be a step, not "This skill is designed to...".
• Vague trigger conditions. "Use when appropriate" is not a trigger. "Use when the user says X or does Y" is.
• Skipping the pressure test.
• Bloating SKILL.md. Keep it under ~5,000 words. Move detailed reference material (API patterns, error catalogs, large templates) into references/.md and link to it from the relevant step. Large SKILL.md files degrade response quality because all of it loads into context once the skill triggers.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gsarig
• Source: gsarig/skills
• License: MIT

Install and usage instructions live in the source repository linked above.