What problem does it solve? High-stakes or uncertain engineering work often fails because agents trust memory on fast-moving APIs, guess decisions that belong to the user, implement inline instead of delegating, or declare done on green tests alone. This Skill enforces a disciplined operating loop that prevents those failure modes. ## Core Features & Use Cases - Ground-truth research: Verifies uncertain claims against primary sources (official docs, actual source code) using parallel read-only agents instead of relying on recall. - Adversarial verification: A second independent pass tries to refute every load-bearing finding, defaulting to the pessimistic reading unless docs are explicit. - Structured decision surfacing: Presents genuine user decisions one at a time, each with a recommended option and stated trade-off. - Delegated implementation with verification: Splits work into disjoint slices for parallel writers, then runs the real gate (build, full tests, actual behavior) before documenting and shipping. - Use Case: When upgrading a fast-moving library or changing shared config with real blast radius, invoke this loop to verify the API against live docs, surface migration choices to the user, delegate the code changes, and confirm behavior before release. ## Quick Start Ask the agent to run the research-decide loop on your task, for example by invoking /research-decide migrate the API client to the new library version.
