skill-factory

skill-factory creates, updates, and standardizes working agent skills. It preserves the requested purpose while producing a portable directory whose instructions, references, assets, scripts, tests, Mise graph, CI, examples, and evals pass fresh checks.

Outcome and motivation

The outcome is a skill that produces the intended user-visible result and can prove it. Package shape is necessary evidence, never the result by itself. State the reason behind every material constraint so the executor can adapt inside the allowed boundary without defeating the outcome. Keep deterministic mechanics out of prose so model attention stays on semantics, judgment, creative work, and exceptions.

Simplicity and language

Keep this factory and every output as the smallest coherent structure that preserves the outcome, proof, boundaries, forbidden outcomes, and mandatory methods, plus accepted behavior. Simple means easy to reason about: one canonical owner per rule, one stable term per concept, one default path per job, and one explicit failure branch for each material unknown. Remove duplicate rules, decorative structure, needless indirection, and choices that do not change behavior. Never hide essential domain complexity or weaken a rule to reduce lines.

Set a context budget for each operation before loading support. Keep the always-loaded body as the smallest domain-complete decision path. At each branch, load only the canonical owner needed for the current decision, record or verify its content digest, reuse current verified receipts, and reread only after owner change or uncertainty. Token efficiency fails if it removes a domain motive, constraint, interface, failure branch, proof duty, model-owned capability, or the context needed for intelligent and thoughtful adaptation.

Use plain and direct language. Load the writing rules through mise run lint-writing for this factory and every skill it creates, updates, standardizes, or imports. Put the result first, use active verbs, keep one idea per sentence and one topic per paragraph, and place instructions in execution order. Mechanical lint finds bounded text faults. A same-meaning human review decides whether the package is simpler and still complete.

Evidence and source order

Use the user's request and live owning sources first. Then use the current target skill, the generation contract through mise run validate, task-specific references, worked examples, and eval evidence. Load use-case specificity through mise run domain-research-policy before defining domain behavior, terms, roles, constraints, or proof. Every aspect and primitive must state its domain role, protected outcome, concrete progress value, motivation, prevented failure, and proof. Load resource and experiment design through mise run improvement-policy before choosing a data structure, format, cache, benchmark, or improvement trial. The required generated-skill order is Outcome, Motivation, Evidence, Mise task graph, Steps, Assets, then Evals. This order links intent to action and action to proof.