chaychun/house-rules
Nine skills for main-thread feature development, delegated work, review, and human testing.
Investigate a bounded architecture, integration, or API question as a read-only subagent for the main agent. Compare current code, official documentation, established patterns, tradeoffs, and risks; report findings without making product decisions or implementing a solution.
Build one bounded visual or interaction design variant as a subagent for a main-thread exploration. Use the agreed real-codebase scratch surface, keep the variant distinct and reviewable, and report its limits without selecting or shipping a winner.
Diagnose and repair an observed bug, test failure, or unexpected result as a bounded subagent assignment. Trace the cause, make a focused fix, check the affected path, and return evidence to the main agent; not for speculative feature work.
Use in the main thread when distinct visual or interaction directions should be made tangible for the human to compare. Frame the question, delegate separate variants with build-variant, and discuss their merits; use before or during design or guided iteration, not for routine UI tweaks.
Implement a bounded outcome as a subagent for a main feature thread. Work within the human-agreed behavior and general architecture, decide implementation details, verify proportionally, and report progress or consequential deviations to the main agent.
Drive discussion and investigate uncertainty for a feature or fix in one durable conversation. Implement after the human approves a direction, or directly when a focused request specifies both the change and its approach. Own guided edits or delegated implementation, oversight, testing, and next steps toward shipping; delegate broad reads to protect the main context.
Inspect implementation choices for consequential shortcuts, loopholes, and brittle workarounds as the main agent or an independent reviewer. Compare with repository conventions and reliable documented paths; report actionable risks, not style preferences.
Check an implementation against behavior actually confirmed by the human, either as the main agent's focused check or an independent review assignment. Find material matches, mismatches, omissions, and unplanned behavior without inventing requirements or changing code.
Run an interactive development manual test session in the main thread after implementation is ready for human exercise. Present one action with its immediate expected observation, collect feedback, and coordinate fixes without delegating the conversation. Not for realistic-data testing in beta, preview, or production.