Skills
problems-firstPressure-test feature requests, product ideas, PRDs, and user stories in the Problem Space before writing specs or code. Peel back surface solution requests with the 5 Whys, audit user workarounds and situational context, filter scope bloat on the Kano curve, distinguish local-maximum tweaks from higher-ceiling workflows, and issue a Stop, Prototype-to-Learn, or Build-to-Earn gate verdict. Use whenever a user asks to scope a feature, write a PRD or user story, validate a startup or product bet, prioritize a backlog, plan an MVP, or jump straight into building a new app or workflow.
shared-understandingTurn messy notes, voice transcripts, Slack threads, and unstructured product ideas into clear, plain-language product briefs ready for engineers or coding agents to decompose. Strip corporate PRD bloat, preserve concrete human context, separate locked decisions from open questions, define hard boundaries and non-goals, and break work into thin vertical slices. Use whenever a user dumps rough thoughts, brainstorms, call notes, or half-formed feature ideas and wants a clear spec, engineering handoff, epic brief, feature one-pager, or agent-ready plan.
architecture-restraintMatch product workflows to the simplest reliable computing layer (deterministic code, classical ML ranking, single-turn structured LLMs, or stateful AI agents) and design clean boundaries for runtime context, caching, tools, and coding harnesses. Use whenever designing system architecture, deciding whether a feature needs an LLM or agent, defining MCP or API tool schemas, managing token budgets and prompt caching, structuring CLAUDE.md or AGENTS.md files, or reviewing technical specs for over-engineering and security risks.
eval-hill-climbBuild trustworthy AI product evaluations from real traces and run disciplined hill-climbing loops without leaking held-out test sets or overfitting rubrics. Guide manual trace open-coding, failure taxonomy clustering, deterministic assertion writing, human-calibrated binary LLM-as-judge design with strict train/dev/held-out-test isolation, and CI release gating. Use whenever evaluating an LLM feature, debugging agent failures, designing eval datasets or LLM judges, running prompt or harness optimization, or setting AI launch gates.