Skill Builder

Build reusable capability from working procedures, domain knowledge, and recurring preferences. Choose instructions, evidence, and tools that make the task succeed.

Quick Path

1. Infer the outcome, trigger, target harness, edit scope, and authority from the request and package. Ask only for consequential missing information.
2. Identify evidence: a real task, artifact, documented contract, preference, or correction. Diagnose failures before choosing what to change.
3. Choose coherent scope. Keep coupled steps together; split unrelated tasks with different triggers, tools, or permissions.
4. Edit the source and affected references together. Put routine instructions in SKILL.md, optional detail in references, repeatable operations in scripts, and output structures in assets.
5. Validate changed surfaces. Report format, execution, discovery, and behavioral evidence separately.
6. Deliver the usable patch, material decisions, and remaining limitations. Publication and installation follow the requested scope.

Read local layout for repository and installed copies. Read source guide for rule provenance and target-specific differences.

Choose What Belongs

Add what the agent cannot reliably infer: a preferred explanation style, private schema, subtle API contract, required artifact, fragile operation, or recurring mistake. Start from real work and user preferences.