Generate a project Skill

Create a compact, searchable Skill from the project itself. This Skill is for maintainers who author a draft Skill they own.

Inputs

Use the project directory or prepared project source. Ask for the destination only when the request does not provide one.

Inspect

1. Read project instructions and decision records first.
2. Read manifests, exported entry points, and build configuration.
3. Find the test, lint, typecheck, build, and release commands.
4. Trace the main public workflows through their real entry points.
5. Note generated directories and files an Agent must not edit.
6. Check recent changes when they explain current conventions.

Project navigation

The Skill must let an Agent find any part of the project without a prior index.

1. Name the project from its manifest. Use the directory name when no manifest names it.
2. List the entry points the manifest declares, such as main, module, types, exports, and bin.
3. List the source and documentation directories an Agent reads most.
4. Use project-relative paths for every file pointer.
5. Never add a generated directory prefix to a project-relative path.
6. Give search commands the Agent can repeat, scoped to those directories.
7. Prefer rg for search. Name the directories to skip in the command.
8. Tell the Agent when project changes require a new Skill run.

Write the search commands so their results point at real project files.

Skip these paths while collecting project files:

• version-control data, such as .git
• dependency directories, such as node_modules and vendor
• generated output, such as dist, build, target, .output, .nuxt, and .next
• caches and reports, such as coverage
• credential files, such as .env and any private key

Skip any file larger than 512 KB. Do not follow symbolic links while collecting project files.

Output

Write one directory whose name matches the Skill name. The directory must contain SKILL.md. Use references/ for architecture details and command guides. Use scripts/ only for reusable automation.

The SKILL.md frontmatter contains name and description. It may also contain license, compatibility, metadata, and allowed-tools. If the project declares a license, copy its identifier into license. Never infer a license. Include compatibility only for specific tool, runtime, network, or product requirements, at most 500 characters. Omit it for general guidance. Do not repeat the package name, tested version, or implementation details. metadata maps string keys to string values. allowed-tools is an experimental space-separated string of pre-approved tools. Use lowercase letters, numbers, and single hyphens in the name. Keep the name at 64 characters or fewer.

Cross-Agent portability

Keep one source Skill for Claude Code, Codex, Gemini CLI, and other compatible Agents. Their discovery paths differ. Keep installation instructions outside the generated Skill. The Agent Skills specification defines the shared format.

• Describe capabilities, such as reading files or running a command, instead of provider-specific tool names.
• Ask for inputs explicitly. Do not require $ARGUMENTS, dynamic context injection, hooks, or a specific subagent API.
• State the project root as the working directory for project commands. Resolve bundled references and scripts from the Skill directory instead.
• Name required runtimes, binaries, network access, and credentials beside the step that needs them.
• If a required capability is unavailable, report the missing capability and stop that dependent step. Never invent evidence or silently skip a required check.

Check one matching task, one unrelated task, and one missing-input task. Use a fresh session for each Agent the user asks to support when those Agents are available. Record the Agent version, model, task, observed output, and any required permission. Separate format validation, discovery, activation, and task completion in the report. If an Agent is unavailable, report that path untested. Valid frontmatter alone does not prove portability.

Quality checks

• Describe when the Skill applies.
• Do not explain the language, framework, or tools. The reader already knows them.
• Use project terms exactly.
• Point to source files instead of copying them.
• Run project commands only when they add useful evidence.
• Record the observed command and outcome.
• Include repeatable search commands for project source and documentation.
• Keep file pointers rooted at the real project directory.
• Separate rules from optional guidance.
• Remove stale, inferred, or duplicated instructions.
• Link each reference from SKILL.md.

For a direct run, show the generated files for user review. Do not replace an existing Skill until the user approves it. Do not claim that the Skill passed Harness checks.