research

Generate a complete, playable, visually beautiful game in one orchestrated multi-agent workflow. The orchestrator writes an immutable contract + an art-direction style bible, fans the build out across parallel implementers (with dedicated world/structure/character/fx/lighting art agents), runs a review → adversarial-verify → fix gauntlet, then actually runs the game, captures screenshots, and iterates against an art-director judge until it looks great. Stack-agnostic; browser/canvas games are the sweet spot. Composes the contract-first method with front-loaded art direction and a closed screenshot-judge feedback loop.

Test-driven feature-build workflow for an unfamiliar codebase: map it into a navigable wiki and a C4 architecture model, cut the task into vertical walking-skeleton slices (each E2E-testable), drive every slice through /red → /green → /refactor with outside-in TDD, shift-left gating, and atomic feature-scoped commits, then /presubmit (full real gate + parallel multi-agent grading panel). Stack-agnostic.

C4-model documentation tracker with SHA propagation and staleness detection

Research-backed implementation planning with a mandatory clarification dialog. Every open question (undefined terms, ambiguities, scope edges) must be resolved by user Q&A or codebase research before any plan body is written. Produces phased plans at docs/plans/ with SHA-tracked references and an embedded resolved-questions audit table.

Proof-based bug investigation. Forces explicit hypotheses, evidence-based verification, 5 Whys root cause analysis, and falsification check, with a written investigation log at docs/investigations/ before any code change is permitted.

Analyze and address pull request review comments. Fetches unresolved threads via GitHub GraphQL, evaluates each comment against real code, makes changes or drafts responses.

Random plugins I built for myself

Install an opinionated Claude Code statusLine into ~/.claude/settings.json. Layout: <folder> <model> effort: <level> ctx: <human-tokens> / <pct>%. Conditional segments, 1M / 200K context window fallback by model name, non-destructive merge, JSON validation, and dry-run smoke test.