Skills

Not Rest Inc. · Capabilities · Expertise, packaged and versioned

A prompt is advice you give a model once. A skill is expertise the model can load: a written, versioned procedure — steps, rules, output shapes, honesty requirements — that turns a general model into a specialist for one job, repeatably. The difference matters because repeatability is what you can build on. A good answer is luck until the procedure that produced it is written down.

You can read ours

The not.rest plugin — public at github.com/notrestai/notrest, MIT-licensed — ships thirty-one skills as plain, inspectable files. Among them:

• researcher — multi-pass research with tiered real sources and a disconfirmation pass before any recommendation.
• critic — adversarial review: steelman first, then attack, then a severity-ranked verdict.
• factcheck — claim-by-claim verification with honest labels: confirmed, plausible, refuted, misleading, unverifiable.
• decider — decisions structured as weighted criteria, evidence, sensitivity, and a pre-mortem.
• stepbystep → actionplan — a goal turned into a stress-tested plan, then a runbook of exact commands.
• sessionend — the session's state captured into files a successor can resume from.

Three principles, stated and checkable

Token-lean — progressive disclosure: load only what the task needs. Verified — real sources, disconfirmation, honesty labels on every claim. Model-agnostic — the discipline lives in the skill, not in any one model.

That last one is the deep point: skills make capability portable. A model will be replaced; a written procedure survives its executor — and gets better with every revision, on the record.