Build agent skills from a reviewable spec, evaluate their behavior, and improve them over time.

A Skillet skill has three parts:

• spec.md defines the skill's intent and required behaviors.
• SKILL.md gives the agent its instructions.
• evals/cases/*.yaml tests those behaviors in fresh workspaces.

Skillet scaffolds, validates, and evaluates those files. It never calls a model API or handles API keys; it invokes your existing agent CLI for authoring and evals.

Skillet requires Node.js 20 or newer.

npx -y @sentry/skillet@latest init

With pnpm, use pnpx @sentry/skillet@latest init instead. The explicit @latest keeps agent-driven authoring on the current CLI and instructions.

init installs the skillet-authoring skill in user scope so your agent knows how to use the CLI. It uses dotagents and asks before writing under ~/.agents.

If you prefer a global binary, install it explicitly:

npm install -g @sentry/skillet
skillet init

Installed binaries check npm at most once per hour and suggest the current npx command when an update is available. If your agent reads skills from somewhere else, copy the skills/skillet-authoring directory directly into that location.

To install the authoring skill with dotagents directly:

npx -y @sentry/dotagents@latest --user add getsentry/skillet skillet-authoring

add records and installs the skill immediately. Run npx -y @sentry/dotagents@latest --user install later to refresh it.

Or ask your agent to install the skillet-authoring skill for you.

Ask your agent:

Create a skill that enforces our commit conventions.

The authoring skill handles the workflow: scaffold the files, clarify the behavior, write the spec, render the agent instructions, add eval cases, validate everything, and run the evals.

To start manually instead:

npx -y @sentry/skillet@latest new commit-conventions
cd commit-conventions
npx -y @sentry/skillet@latest status

skillet status reads the files on disk and tells you the next step. When writing an artifact yourself, use skillet instructions spec, skillet instructions skill, or skillet instructions evals for its current format and rules.

For an existing skill, run skillet status <path>. Uppercase SPEC.md and structurally invalid lowercase spec.md are treated as migration input; preserve or rename the legacy content, then derive a valid lowercase spec.md before adding eval coverage. Inventory the old skill's triggers, workflow, exact lists and protocols, thresholds, stop rules, constraints, and runtime references first, then reconcile that inventory against the new spec and rendered skill instead of assuming a shorter rewrite is equivalent.

npx -y @sentry/skillet@latest validate
npx -y @sentry/skillet@latest eval --dry
npx -y @sentry/skillet@latest eval --baseline

• validate checks the spec grammar, SKILL.md frontmatter, eval schemas, and behavior coverage.
• eval --dry finds cases that a do-nothing agent would pass.
• eval --baseline runs each case with and without the skill and reports the difference as lift.

Eval runs invoke real agent CLI sessions. They use that CLI's configured account and model.

Example output:

Behaviors:
  conventional-subject: 100% (3/3) | baseline 33% | lift +67%
  branch-safety:        100% (3/3) | baseline 0%  | lift +100%

Lift answers a concrete question: did this skill improve the agent's behavior? Zero lift is useful too—it means the configured agent already passed without the skill.

A behavior in spec.md:

### Behavior: Conventional subject

The agent SHALL write commit subjects as `<type>(<scope>): <description>`.

#### Scenario: Committing a staged bug fix

- **WHEN** the workspace has a staged bug fix and the user asks to commit
- **THEN** the commit subject starts with `fix` and stays under 70 characters

One case that covers it:

behavior: conventional-subject
prompt: |
  I fixed the null check in app.js—please commit my staged change.
setup: |
  git init -q -b main
  git add -A
checks:
  - shell: git log -1 --format=%s | grep -Eq '^(feat|fix|chore)'
  - judge: The commit message accurately describes the null-check fix.

Checks inspect the resulting workspace, not just the agent's response. Use file_exists and shell for deterministic checks; use judge when the requirement needs semantic evaluation. See examples/ for a small skill plus full Garfield and Effect conversions with preserved upstream snapshots.

Every command supports --json. Agent-driven workflows use npx -y @sentry/skillet@latest <command> (or the pnpx equivalent); the table uses skillet as shorthand. Run skillet <command> --help for command-specific options.

Skillet uses Codex by default and has built-in support for Claude Code. Select one with --harness codex or --harness claude, optionally with a model suffix such as --harness claude:sonnet. You can configure another CLI in .skillet.yaml.

Skillet leaves the model unset unless you choose one, so the agent CLI's configured model applies. Built-in trials and judges default to low effort, and up to four trials or skill/baseline variants within a case run in parallel. Override these with --effort, --concurrency, or .skillet.yaml. Baseline remains opt-in.

By default, eval agents run directly on your machine with full access. Use this only for skills and evals you trust. For untrusted skills or CI, build the included Docker image and run with --sandbox docker.

See LIFECYCLE.md for the artifact flow, eval execution model, harness configuration, and sandbox details.