skill-x

Use this skill when the task involves coding-agent interpretability, hidden states, residual streams, linear probes, latent programming horizons, edit-outcome prediction, or monitor-and-steer workflows for coding agents. Applies the linear-probe procedure (Silva, Tu, Monperrus 2026) with expected AUC numbers, horizon k≈25, mid-layer inverted-U, and cross-benchmark transfer. Trigger when the user says "skill zero", "interpret the coding agent", "probe the agent's hidden states", or asks to predict/monitor a coding agent's edit outcomes. Do NOT trigger for general non-trivial tasks without an interpretability angle — use `think-like-gpt-5-6` or `super-skill` for those instead. This skill also contains a legacy General-mode rigor framework in its body for backward compatibility; new callers should prefer `think-like-gpt-5-6` for that use case.

Use when running an expensive frontier model on codebase-heavy or token-heavy work and the user wants that model to orchestrate, architect, synthesize, and final-judge while cheaper helper agents do bounded heavy lifting (large repo scans, long log reduction, narrow code patches, browser/test verification, etc.). Trigger on requests like "delegate this efficiently", "save tokens on this", "use cheap subagents for the heavy parts", or any task where the expensive model's judgment matters but its throughput does not. Treat subagent reports as leads, not facts — always vet before shipping. Do NOT trigger for trivial single-step work, for high-stakes single-source-of-truth tasks (medical, legal, financial advice), or when no cheap subagents are available.

Apply the combined rigor + delegation framework for complex tasks. Two operating modes: (1) **Rigor mode** — use the GPT-5.6 Sol 7-step loop, 10 principles, 9-stage architecture, task classification, risk register, validation layers, acceptance/rejection criteria, adversarial review, and pre-delivery checklist (default for any non-trivial task); (2) **Delegation mode** — for codebase-heavy or token-heavy work that can be split into independent parallel slices, use the Efficient Fable delegation pattern: cheap subagents do the heavy lifting, the orchestrator keeps judgment, architecture, synthesis, and final review; always apply the vetting protocol (treat subagent reports as leads, not facts). Trigger when the user says "super skill", "use the super framework", "delegate this efficiently", or any complex task where you want both auditable artifacts AND cost/speed efficiency. Do NOT trigger for trivial single-step questions, or for tasks that don't tolerate any subagent delegation (legal, medical, single-sourc

Apply the GPT-5.6 Sol execution-intelligence framework to any complex task. Use when the user asks the agent to "think like GPT-5.6", "use the blueprint", "apply the execution framework", or wants auditable, requirement-driven, evidence-aware task execution. Also trigger on requests that benefit from a controlled 7-step loop (Observe → Interpret → Decide → Act → Verify → Repair → Record) with explicit acceptance criteria, validation gates, risk register, and a final pre-delivery checklist. Best fit for: multi-step coding, research with citations, artifact creation, operational actions, high-stakes guidance, and any task where the user wants inspectable reasoning rather than a fluent black-box answer. Do NOT trigger for trivial single-step requests where the loop overhead exceeds value, or when the user explicitly wants a fast informal answer.