ap7x42a

Use when designing, reviewing, or changing Bayesian confidence, calibration, Beta-Binomial updating, prior specification, posterior computation, Thompson/LCB ranking, Brier/ECE calibration, alpha/beta parameters, uncertainty-aware scoring, or probabilistic feedback loops. Also use when a system proposes hardcoded confidence values, magic thresholds, or point estimates where distributions should carry the evidence.

Upgrade a prose-only or weakly tooled Agent Skill into a harnessed skill with reusable scripts, on-demand references, assets, strict static validation, a drift manifest, and credible regression tests. Use when a skill repeatedly re-derives deterministic code, embeds large reference material or templates in SKILL.md, has dangling resource wiring, lacks meaningful tests or receipts, or needs to be made more repeatable and package-ready. This is for upgrading an existing skill; use a general skill-authoring workflow to create one from scratch.

Use structured multi-branch deliberation for complex, ambiguous, high-impact, or failure-prone problems. Create independent solution branches, causal hypotheses, assumptions, evidence, experiments, critiques, and decision criteria; prune invalid branches, merge compatible discoveries, red-team leading candidates, and produce a traceable synthesis. Preserve active branch state across interruption or compaction. Skip simple deterministic tasks.

A procedure for evaluating or improving a contestable artifact — a methodology, prompt, skill, spec, architecture proposal, or "does this approach actually work" claim — by building a falsifiable test instead of offering an opinion. Use it whenever you're asked "is this any good / will this work / what do you think of this / can you improve this" about something whose quality is arguable and where a plausible-sounding answer would be cheap and unverifiable. Skip it for simple factual lookups or matters of pure taste with no testable claim.

Use when a problem benefits from nonstandard reasoning lenses: abductive, negative-space, predictive, asymmetric-belief, external-memory-binding, concurrent, split-level, or stigmergic thinking. Apply these as structural analysis prompts, not as magical tools. Useful for difficult debugging, strategy, research synthesis, postmortems, architecture, and reviewing reasoning for gaps or hidden assumptions.

Use for consequential coding or focused multi-step builds where a plausible-but-wrong result is costly: ambiguous debugging, preserving behavior through refactors, extraction, mutating automation, concurrency, external dependencies, stalled projects, scope drift, or release decisions. Also use when tempted to report a diagnosis, fix, or completion without an external check that could return "no". Skip trivial one-pass edits.

Use before confirming OR refuting any claim about system state ("is X in the data?", "did this stop?", "are those still there?"). Enumerate every store the evidence could live in and query each; a partial sweep can only report "absent from what I checked," never "refuted" — and the store you skip is usually the first-class one you don't know how to open.

Use when the current agent is asked to iteratively improve code, prompts, tests, or skill behavior in the same loaded session without spawning child Claude or Codex sessions.

Use before drawing any conclusion from an accumulating operational store — a DB table, telemetry/event stream, JSONL ledger, or metric that grows over time. Anchor the analysis to the events that changed the data-generating process; a plain aggregate over a non-stationary store will invert your causality.

Use when working an inherited system, artifact, claim, metric, test, or prior agent result you did not author and cannot fully trust -- especially long-lived AI-built systems where intent and implementation have drifted. Read every load-bearing artifact as a possible cursed wish: something that satisfies the literal request while betraying the intended outcome. Also use before trusting, extending, deleting, or citing any inherited behavior, and when reviewing a system's own account of how it did. The complement to building (fable-method) and proving a claim (evaluate-by-experiment): how to stand toward what you inherited.

Preserve visible operational continuity across context compaction, session resume, model handoff, and long-running agent work. Capture exact user inputs, directives, corrections, preferences, task state, execution cursor, brainstorming alternatives, decisions, commitments, open loops, evidence, artifacts, and active deliberation pointers in a transactional, provenance-backed ledger. Use when Claude Code or Codex must resume work without asking the user to reconstruct prior context.

Use at the start of substantial agent sessions that need coordination, delegation, adversarial review, handoff discipline, or survival across context clearing. The conductor keeps judgment, defines contracts, delegates bounded work only when the current host exposes delegates, verifies receipts, and writes a human-readable handoff at any human-needed breakpoint.