New Skill

Creating a SKILL.md

A skill is a markdown file that teaches an AI agent a specific workflow, tool, or pattern. Skills are loaded into context when the agent recognizes a task that matches the skill's description.

File location

Place the skill in a top-level skills/ folder at the repository root:

skills/<skill-name>/SKILL.md

For example: skills/critique/SKILL.md, skills/errore/SKILL.md.

Do not put skills inside package folders like cli/skills/, website/skills/, or packages/foo/skills/ unless the repository intentionally syncs or mirrors them there for internal tooling. The canonical repository layout for a skill you are creating is always the root-level skills/ directory.

The folder name should match the skill name in kebab-case. Each skill gets its own folder so it can include companion files if needed (scripts, templates, references).

For personal skills that follow you across all repos and are not meant for distribution in a GitHub repository, place them in:

~/.config/opencode/skills/<skill-name>/SKILL.md

Personal skills are only available on your machine. Repository skills are shared with everyone who clones the repo.

Editing skills synced from other repositories

Some projects (like kimaki) sync skills from external GitHub repositories into a local skills folder. If a skill was synced from another repo, never edit the synced copy. The synced folder is overwritten on every sync and your changes will be lost.

Instead, find the source repository where the skill originates and edit the SKILL.md there. The sync process will pick up the changes on the next run. If you are unsure which repo a skill comes from, check for a sync script (e.g. scripts/sync-skills.ts) or a source-repo field in the skill's frontmatter.

Distribution and installation

When you publish skills in a GitHub repository, other users can install them with the skills CLI:

npx skills add owner/repo

This downloads the skills from the repo and symlinks them into the user's agent directories. Add this to your repo's README so users know how to install:

## Install skill for AI agents

\`\`\`bash
npx -y skills add owner/repo
\`\`\`

This installs [skills](https://skills.sh) for AI coding agents like
Claude Code, Cursor, Windsurf, and others. Skills teach agents the
workflows, patterns, and tools specific to this project.

Making a skill installable from a URL

Use this when the project does not have a public GitHub repo (private repos, closed-source SaaS products) or when you want the install command to use your own domain:

npx skills add https://example.com

The skills CLI treats any non-GitHub/GitLab HTTP(S) URL as a well-known discovery endpoint (RFC 8615). It fetches /.well-known/agent-skills/index.json and expects the v0.2.0 discovery schema:

{
  "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
  "skills": [
    {
      "name": "my-skill",
      "type": "skill-md",
      "description": "One-line description shown in the CLI skill picker.",
      "url": "./my-skill/SKILL.md",
      "digest": "sha256:<hex sha256 of the exact artifact bytes>"
    }
  ]
}

Rules:

• The digest must match the served bytes exactly. The CLI downloads the artifact from url and verifies its sha256 before installing. Never hardcode the digest; compute it from the same string you serve.
• url is resolved relative to index.json, so ./my-skill/SKILL.md works and survives domain changes. Absolute URLs also work.
• name must be kebab-case (^[a-z0-9-]+$, no leading/trailing/double dashes) and match the SKILL.md frontmatter name.
• type: "skill-md" serves a single markdown file. Use type: "archive" (zip or tar.gz with a root SKILL.md) only when the skill ships companion files.
• Keep the canonical SKILL.md in the repo (skills/<name>/SKILL.md) and serve those exact bytes, so the published skill can never drift from git.

Example with spiceflow on Cloudflare Workers: import the SKILL.md with vite ?raw and compute the digest once per isolate with WebCrypto.

import { Spiceflow } from 'spiceflow'
import skillMd from '../../skills/my-skill/SKILL.md?raw'

let digestPromise: Promise<string> | undefined
function getSkillDigest(): Promise<string> {
  digestPromise ??= (async () => {
    const bytes = new TextEncoder().encode(skillMd)
    const hash = await crypto.subtle.digest('SHA-256', bytes)
    const hex = Array.from(new Uint8Array(hash), (b) => b.toString(16).padStart(2, '0')).join('')
    return `sha256:${hex}`
  })()
  return digestPromise
}

export const skillDiscoveryApp = new Spiceflow()
  .route({
    method: 'GET',
    path: '/.well-known/agent-skills/index.json',
    handler: async () =>
      Response.json({
        $schema: 'https://schemas.agentskills.io/discovery/0.2.0/schema.json',
        skills: [
          {
            name: 'my-skill',
            type: 'skill-md',
            description: 'One-line description shown in the CLI skill picker.',
            url: './my-skill/SKILL.md',
            digest: await getSkillDigest(),
          },
        ],
      }),
  })
  .route({
    method: 'GET',
    path: '/.well-known/agent-skills/my-skill/SKILL.md',
    handler: () =>
      new Response(skillMd, {
        headers: { 'content-type': 'text/markdown; charset=utf-8' },
      }),
  })

Validate end-to-end against a local dev server before deploying:

npx skills add http://localhost:3000 --list
npx skills add http://localhost:3000 -y --copy

If the digest is wrong the CLI silently skips the skill, so an empty --list result usually means a digest mismatch.

Frontmatter

Every SKILL.md starts with YAML frontmatter containing three required fields:

---
name: skill-name
repo: remorses/critique # owner/repo format, matching the GitHub repository
description: >
  One to three sentences explaining what this skill does and when to use it.
  Start with a noun or verb phrase. Include trigger conditions so the agent
  knows when to load this skill automatically.
---

• name: kebab-case identifier matching the folder name
• repo: the GitHub repository that owns this skill in owner/repo format (e.g. remorses/critique). This lets users reinstall the skill later with npx skills add owner/repo. Always use the canonical source repository, not a synced copy or generated package path.
• description: this is the most important field. The agent reads descriptions of all available skills and decides which to load based on this text. Be specific about when the skill applies. Include keywords the user might say.

Good description example:

name: critique
repo: remorses/critique
description: >
  Git diff viewer. Renders diffs as web pages, images, and PDFs
  with syntax highlighting. Use this skill when working with critique
  for showing diffs, generating diff URLs, or selective hunk staging.

Bad description example:

description: A helpful tool for developers.

File structure

After the frontmatter, write the skill as a normal markdown document. Follow this general structure:

# Skill Title

One paragraph explaining what this skill is and why it exists.

## Key section

Core rules, commands, or patterns. Use code blocks for commands
and examples. Use numbered lists for sequential steps.

## Another section

More detail, edge cases, gotchas, tips.

There is no rigid template. Structure the content in whatever way communicates the workflow most clearly. Some skills are short (20 lines for a simple CLI tool), others are long (600+ lines for a complex pattern like errore).

Writing style

Write for an AI agent, not a human. The reader is a language model that will follow these instructions while helping a user. This changes how you write:

• Be direct and imperative. Say "Always run tool --help first" not "You might want to consider running the help command."
• Include concrete commands and code. The agent needs copy-pasteable examples, not abstract descriptions.
• State rules as rules. Use "Never", "Always", "Must" when something is non-negotiable.
• Show the right way, not just the wrong way. After saying what not to do, immediately show what to do instead.
• Use code blocks with language hints. The agent uses these to generate correct code.
• Keep prose short between code blocks. One or two sentences of explanation, then an example.
• Call out common mistakes. If there is a gotcha the agent will likely hit, warn about it explicitly.
• Do not add HTML comments. Skills are instructions, not generated files. Avoid comments like <!-- Skill instructions for agents using ... -->.

What makes a good skill

A good skill captures hard-won knowledge that is not obvious from reading docs or source code alone. Focus on:

• Correct usage patterns — the commands and code that actually work, not just what the docs say
• Gotchas and edge cases — things that break in subtle ways (e.g. "libsql transaction() with file::memory: silently uses a separate empty database unless you add ?cache=shared")
• Opinionated defaults — when there are multiple ways to do something, state which way to use and why
• Integration context — how this tool fits into the broader workflow (e.g. "Always use critique when showing diffs to Discord users because they cannot see terminal output")

A bad skill is just a copy of the tool's README or man page. If the agent could figure it out from --help, it does not need a skill for it.

Keep the SKILL.md thin — point at canonical docs

The best skills are thin. Keep as much user-facing documentation as possible in the repository docs, README, or CLI help output. The SKILL.md should point agents at those canonical docs so content stays fresh, then add only the agent-specific instructions that do not belong in human-facing docs.

Good agent-only instructions include rules a human reader does not need, but an agent must follow to avoid mistakes:

• When to load the skill and which keywords should trigger it
• Commands to run first, like fetching a README or reading full --help output
• Never truncate rules for documentation commands
• Agent workflow constraints, such as which files not to edit, which package manager to use, or how to validate changes
• Common model failure modes, like using stale APIs, editing generated files, or skipping required setup

Do not copy the whole README into the skill. Put general explanations, examples, API docs, and user-facing guides in the repo docs instead. The skill should be the small bridge between the agent and the canonical source of truth.

There are two variants:

1. CLI tools → run <tool> --help

Put as much documentation as possible into the CLI itself — command descriptions, option help text, examples. The skill then says:

Every time you use mytool, you MUST run:

\`\`\`bash
mytool --help # NEVER pipe to head/tail, read the full output
\`\`\`

Exception: some CLIs have a dedicated <tool> skill subcommand when --help is not rich enough (e.g. playwriter skill). Prefer --help by default and only use a custom subcommand when the CLI ships one.

2. Libraries and projects → curl the raw README

For libraries, frameworks, and pattern skills, keep the canonical docs in README.md and have the skill curl the raw file from the main branch so the agent always reads the latest version:

Every time you work with myproject, you MUST fetch the latest README:

\`\`\`bash
curl -s https://raw.githubusercontent.com/owner/repo/main/README.md # NEVER pipe to head/tail
\`\`\`

In monorepos/workspaces, always put the README at the repository root — not inside individual package folders. Package-level READMEs don't get read by anyone. One root README is the single source of truth for the whole project. The skill should curl the root README path (.../main/README.md), not a package subdirectory.

Never truncate docs output. The agent must read --help and curl'd README output in full. Never pipe through head, tail, sed -n, awk, | less, or any command that strips or limits lines. Critical rules are spread throughout the doc, not just at the top. Agents truncate frequently and miss important context — forbid it explicitly in the skill body.

Examples from real skills

Simple CLI tool skill (critique):

---
name: critique
repo: remorses/critique
description: Git diff viewer. Renders diffs as web pages, images, and PDFs
  with syntax highlighting. Use this skill when working with critique for
  showing diffs, generating diff URLs, or selective hunk staging.
---

# critique

Git diff viewer that creates shareable web pages, images, and PDFs.

Always run `critique --help` first. The help output has all commands,
options, and examples.

## Show a diff

\`\`\`bash
critique --web "Describe pending changes"
\`\`\`

Pattern/convention skill (errore — 647 lines):

---
name: errore
repo: remorses/errore
description: >
  errore is Go-style error handling for TypeScript: return errors instead
  of throwing them. ALWAYS read this skill when a repo uses the errore
  "errors as values" convention.
---

# errore

Go-style error handling for TypeScript. Functions return errors instead
of throwing them.

## Rules

1. Always `import * as errore from 'errore'` — namespace import
2. Never throw for expected failures — return errors as values
3. Use `createTaggedError` for domain errors
...

Notice both follow the same pattern: minimal frontmatter, clear title, actionable content with code examples. The simple tool skill is short and focused on commands. The pattern skill is long and focused on rules and conventions.

Checklist

Before saving a new skill:

1. Does the description clearly state when to load this skill? Would an agent reading just the description know whether to load it?
2. Does the name match the folder name?
3. Does the repo field use owner/repo format pointing to the canonical GitHub repository?
4. Does the skill point at a single source of truth (README curl URL or --help command) instead of duplicating docs inline?
5. Is there an explicit "never truncate" rule next to any docs command?
6. Are there concrete code examples for the main workflows?
7. Did you capture the gotchas — the things that took trial and error to figure out?