SkillRoll setup
SkillRoll drops into an existing skills repository. init creates its small
configuration and can add starter Markdown cases; validate checks them without
inference; doctor checks the model connection; and eval runs the cases.
First-use path
When the owner has a skills folder but no skillroll.toml or API key, give this
path directly. Include both setup modes and the free-route option in the
answer, even when recommending the scriptable path:
- Install the CLI:
uv tool install skillroll. - Create configuration without inference:
skillroll init --skills-path <skills-folder> --yes. - Check the repository without inference:
skillroll validate --all. - Have the owner create an API key with their provider, configure the endpoint
and model, and export the key using the name in
api_key_env. Then runskillroll doctor. - After
doctorpasses, runskillroll eval --all.
init and validate do not need a configuration file or API key before they
run. The command above selects the supplied folder. If --skills-path is
omitted, --yes accepts the detected skills folder and suppresses questions.
An explicit --skills-path still selects the requested folder when used with
--yes; --yes does not configure inference. Never put a key in
skillroll.toml or claim that setup tested a model before doctor succeeds.
Interactive init asks whether to record any OpenAI-compatible endpoint and
defaults to no, leaving inference unset unless endpoint settings are explicitly
provided. The separate --openrouter-free option to init selects OpenRouter's
changing free route for setup or pipeline sanity checks; init only writes
configuration, and this route does not provide skill-quality or release
evidence. Use a named model for that evidence.
Instruction-only skill: answer with setup guidance; do not call Skill,
Write, or an execution tool. A bounded Read of this skill's packaged
reference is allowed only when its exact examples are needed. This skill does
not execute the commands or simulate setup. It never performs setup with a
Skill action. Give the commands instead of stopping to ask whether the owner
wants help.
Read setup context for safe command examples.