| name | skill-generator |
| description | Use when creating, revising, or auditing an agent skill: model its task, boundary, and verification before drafting. |
Skill Generator
Select the verification contract first
At the outset of every create, revise, or audit task, enumerate installed TOML verification presets,
filter them against the task and available runtime capabilities, and show their paths and material
constraints. Recommend one compatible preset with reasons, but never select it. Require the human to
select that preset or provide a complete ad-hoc TOML strategy, then record its path and SHA-256 in
the task model before further modelling or drafting.
The selected strategy records positive triggers, hard near-neighbour non-triggers, task-behaviour
cases, isolated prompt paths, absolute acceptance, deterministic validators, evidence classes, and
candidate/baseline arms. Baselines are either no-skill or an immutable old-skill reference; never
mutate a baseline while testing its candidate. Run candidate and baseline cases in the same
evaluation window.
Classify every result as mechanical, sampled_behavior, comparative, model_judgement,
human_judgement, or unverified. Never promote sampled or judged evidence to a deterministic
claim. Reject strategies that leak diagnoses, rubrics, or expected answers into prompts.
Model the job
Name the invocation as create, revise, or audit in the recorded task model. Derive jobs, variation,
requester-owned choices, and activation boundary from the request, examples, and repository facts.
Route standing conduct to the agent contract, guaranteed rules to tooling, and repository knowledge
to documentation. Show the model, routing, assumptions, and selected verification strategy before
writing or auditing.
Write and verify
Put the activation boundary and neighbours in description. Write instructions as actions taken at
the point they matter. Add resource directories only when the body names their load condition. Use
lowercase letters, digits, and hyphens for the folder and run every shipped script.
Execute the selected isolated cases. Enforce absolute acceptance before comparative deltas: a
candidate that beats its baseline but misses the contract still fails. Aggregate candidate against
its simultaneous no-skill or immutable-old-skill baseline, preserve raw run references, and report
divergence without optimizer, viewer, or automated grader substitutions.