maintain-verification-skill-pi

Draft or revise a named user's Pi -mode skill from their approved preferences and bounded evidence.

Compare separately produced candidate solutions to one approved non-trivial Pi task, cross-judge against a rubric, select a base, graft only the best parts…

Vet an approved performance result before reporting or acting on it: limiter, tuning, physical limits, errors, variation, user relevance and proof that…

Investigate what an approved change could break beyond its diff, find the one or two safety facts and verify them against real code when authorized.

Sketch caller-first types, signatures and module boundaries for an approved non-trivial Pi engineering change before implementation; compare distinct designs…

Design and run a bounded, evidence-linked Pi playbook for an approved complex task when no narrower method fits.

Explain how an approved subsystem works and where behavior belongs using traced entry points, ownership and runtime boundaries.

Adversarially challenge an exact approved Pi code or design artifact with independent evidence-based reviewers, then triage findings without auto-applying them.

Challenge comments and suppressions in an approved diff, prove which are justified, and propose structural fixes before any scoped deletion.

For an approved non-trivial Pi task, select a source-mapped playbook, work in bounded verifiable units, cite loaded principles and stop at owner-held effects.

Design an owner-approved local page whose server sends a small JSON action to a bot webhook, with the sender secret kept server-side and tailnet exposure…

Find a recurring, evidenced agent mistake in an approved repository and prevent it structurally, then prove the guard catches a real past failure.