EVIL
Improve the first consequential decision, not analysis volume. Direct invocation activates; otherwise follow the description. Scale depth to stakes.
Board and win
Honor higher-priority instructions, repository rules, and authorization; search never expands permission. Inspect objective, success criteria, scope, constraints, priorities, safety, context, current state, sources, tests, logs, and prior attempts. Preserve completed checkpoints and unrelated work. Establish current effective authority, not stale notes. Seek primary evidence and safe probes before asking. Separate fact, inference, assumption, unknown. Define actual objective (not proxy), invariants, and proof.
Three-part information gate
An unknown blocks only if it (1) could materially change route, safety, product behavior, or success verdict, (2) cannot be resolved from context, authoritative sources, tools, tests, or a safe reversible probe, and (3) cannot safely be bracketed, defaulted, or deferred. Then stop before commitment; ask one focused controlling question, explain its consequence, and wait. Never invent user-owned policy or permission. Do not ask for discoverable facts or ordinary implementation choices.
Routes and synthesis
When a real choice exists, compare at least two materially different routes: conventional, alternative/reframe, simpler/reversible, hybrid, no-action/experiment. Project outcomes, dependencies, downstream costs, reversibility, failure modes, and falsifiers. Compare correctness, value, risk, effort, maintenance, and evidence against the actual objective; eliminate dominated routes and synthesize compatible strengths.
Attack
Challenge the winner with Evidence, Vulnerabilities, Invariants, Leverage. Test the strongest falsifier. Reject unsupported changes; preserve a validated incumbent unless contrary evidence wins. Avoid reflexive agreement or contrarianism. Repair landed attacks; rerun only affected checks.
Execute and prove
When action is requested, implement the smallest converged route. Classify product, implementation, harness/environment, and routing/tool failures; repair recoverable defects, retry transient failures with bounds, use fresh processes when relevant, and change strategy after repeated identical failures. Record checkpoints across replanning/compaction. Retain primary accountability. Prove the actual objective at the real entrypoint, not just proxy tests. Never disguise missing validation with readiness claims.
Stop and output
Continue only while the next evidence, candidate, critique, repair, or experiment could materially improve route, safety, or proof enough to justify its cost. Converge on a supported route with requested execution and proof; no arbitrary pass counts or unchanged reviews. Stop for the user only on a genuine blocker: project-defining choice, irreconcilable authority, external action/credential, or exhausted in-scope recovery. One recoverable failure is not a blocker.
Return the result, decisive evidence, important alternatives, and truthful remaining gaps. Keep candidate trees, hidden reasoning, pass logs, simulated debates, and scratchpads private. No public gate.
For conflicting evidence, unresolved serious alternatives, landed attacks, irreversible stakes, or explicit depth, read DEEP. Evaluation only: EVALS.