Cross-Skill Evidence Distillation
Promote only recurring actionable principles with a stated failure consequence and a comparison against existing rules.
These examples and illustrative results are independently authored teaching materials, not measured model results.
Use case
Three skills repeat source retention while another has one vendor command. Distillation selects transferable actions, not every heading as policy.
Mechanism
Read candidates and full existing rules, recording repeated sections, actions and failure risks. Compare semantics and propose append/revise/already-covered/too-specific-local rather than new wording as new meaning. Preserve exceptions/review authority; proposals do not automatically update global memory or permissions.
Bad example
Copy every heading globally, duplicate old meanings and require vendor flags everywhere.
Good example
Propose source/observation checks with repeated locations and stale-use risk. Mark sufficient existing coverage Already Covered or propose Revise for gaps. Keep vendor recipes local with scope/exceptions instead of automatic global updates.
Why the change matters
Recurrence/actionability identify candidates and semantic comparison limits maintenance. Failure risks make abstraction operational rather than vague.
Observable expectation
Teaching tables contain source/action/risk/existing_rule/verdict. Covered repeated meaning adds nothing; single CLI detail stays local. Heading counts establish neither correctness nor universal fit.
Limits
Sources can copy one another; recurrence is not truth. Verify facts/applicability. Two-source filters are not statistical evidence, and distillation grants no memory-writing/source-execution authority.