Agent skills

tunables ships three skills for coding agents. Each is a self-contained SKILL.md with its references, written to the same principles as the library, so an agent working in your app holds the same line a careful person would: one home per value, no fallback defaults, every change a revision, access left to your app.

Installing them

npx tunables install

It installs the skills for Claude Code, Codex CLI, Cursor, Gemini CLI and GitHub Copilot, in the project or for your user. See Install skills. After upgrading tunables, run npx tunables update so the skills match the library.

In Claude Code you can also add the repository as a plugin marketplace:

/plugin marketplace add itai-gendler/tunables
/plugin install tunables@tunables

Using them

Ask in your own words; each skill's description tells the agent when it applies. For example:

• "Adopt tunables in this app." runs tunables-adopt.
• "Make the retry count tunable." runs tunables-define.
• "Let each workspace lower its AI suggestions a day, never above its plan." runs tunables-define, which declares the field's override.
• "Check our settings, report only." runs tunables-verify without changing anything.

Or name the skill: /tunables-define add a setting for the export batch size.

What they share

• They propose before they write. Adopt and define present what they would change, including every choice you did not state, and wait for your approval. Verify reports every finding first, and fixes only the ones you choose.
• They use the CLI. Each skill runs the tunables CLI through its own runner, at the version it was installed with, so its checks are the library's own. check, env, explain and model draft need no database; the skills run sync, set or restore only as part of a change you approved.
• They never write tunables' tables directly. Values move through saves and tunables set, with a reason, so the history records each move; a workspace's or a person's own value moves at its scope, with --scope.
• They never decide product meaning in code. What a setting is, who tunes it and what it starts at are product decisions. In a repository with a pdd-kit Product Model, they draft the facts with tunables model draft and bring them to /pdd-ideate; they never write product-model/ themselves.
• They follow your repository. Each reads your agent instructions first, wires into the database client, server, sign-in and UI you already have, and never stages, commits or deploys unless your instructions or you say to.