What problem does it solve? Open architecture and technical-decision questions often get answered with the first plausible idea, with no devil's advocate and no grounding in evidence. This Skill runs a structured research-propose-critique-refine loop that converges on a defensible recommendation instead of a confident guess. ## Core Features & Use Cases - Grounded research: Gathers external prior art via Exa web search and internal evidence from the codebase before drafting anything, with every claim cited. - Separated critic: Spawns a read-only synthesis-critic agent in its own isolated context that scores the proposal against a frozen rubric and returns severity-ranked findings with a PASS/REVISE verdict. - Bounded convergence: Stops on PASS, no progress, oscillation, or a hard cap of 3 iterations, then emits an ADR-shaped decision memo with alternatives, tradeoffs, and open questions. - Use Case: Ask "what's the right cache invalidation strategy for this service" and receive a decision memo comparing 2-3 genuine alternatives, each rejected for a specific cited reason. ## Quick Start Ask the assistant to synthesize the best architecture for your problem, for example: "Research and propose the best wire format for our SDK boundary, iterate until it converges."
