expert-trace-skill-distillationlisted

# Expert Trace Skill Distillation Build skills from evidence, not from a single unconstrained generation pass. The output must be an inspectable software artifact whose sources, boundaries, evaluation results, corrections, and rollback points remain visible. ## When to Activate Activate when: - Distilling a teammate's procedures, review standards, or decision heuristics into a skill - Generating a skill from successful and failed agent trajectories - Turning a repository, manual, API specification, or research paper into reusable procedures - Creating a role-grounded skill without open-ended impersonation - Evaluating whether a generated skill actually helps instead of merely sounding plausible ## Paper-Backed Principles The cited papers support these principles: 1. **Skills are interventions.** Compare the same tasks with and without the candidate skill. 2. **Failures and successes both matter.** Contrast nearby successful and failed trajectories to isolate missing checks, reusable procedures, and regression risks. 3. **Capability and behavior are separate tracks.** Encode work practices separately from bounded style or interaction rules. 4. **Skill generation is a pipeline-level problem.** Evaluate the generator, backbone, source type, executor, and verifier together. 5. **Repository-grounded and document-grounded extraction are different.** Repositories hide operational knowledge across code, config, and scripts. Documents often state rules explicitly but