“Done.” Prove it.

Every request becomes rows you can check, and nothing is done until each row has proof.

• prd-to-proofTurns a request into numbered requirements, then proves every row on the real product.
• request-gateA mechanical gate. It says ok only when every row passed a screenshot scan or a readback it ran itself.
• long-horizon-buildPlans multi-day builds into 15 to 60 minute milestones, each with a validation command.
• iterative-loopOne focused change per round, the same honest check, keep only clear wins.
• proof-principlesThe working rules: verified is 1, unverified is 0. Load it always.

Find what a change breaks outside its own diff before your users do.

• what-could-breakFinds what a change could break outside its diff, then runs real code to prove the one fact that makes it safe.
• hard-reviewA strict review where every claim comes with a command the author can rerun.
• make-a-verifierGenerates a verify skill for your repo that drives the real app the way a user does and captures proof.

Check pages, screens and video the way a person sees them, not the way the markup reads.

• visual-qaOpens the rendered result, judges it like the viewer, and scans screenshots for text drawn over text.
• vision-qcChecks generated graphics at scale with any vision model and gets a strict JSON verdict back.
• mobile-no-zoomStops a web app zooming on phones and proves it in a phone-emulated browser.
• smooth-web-videoDrives the page frame by frame instead of screen recording, and gates every video for stutter.

The bugs that only show up with retries, several workers and real users.

• idempotent-retriesFixes duplicate posts, double charges and messages sent twice after timeouts, retries and restarts.
• single-writer-stateStops workers, processes and agents overwriting each other's files, rows, keys and branches.
• prompt-override-debuggingFixes an LLM feature that ignores the user because its system prompt pushes the other way.
• multilingual-text-debuggingTraces broken non-English text: stray punctuation, invisible direction marks, words cut in half.
• error-to-fix-loopWatches a service log and hands each new code error to an agent that must reproduce it before fixing it.

Build the agents and deciders themselves so they measure, not guess.

• data-kernelAsks questions of big logs and JSON without flooding the agent's context.
• llm-judge-evalsLLM-as-judge evaluations that give the same answer twice.
• autoresearch-loopUnattended optimization of one asset against one number, one change per round, every round logged.
• agent-cli-fallbackCalls claude -p or codex exec from a service, with a fallback provider and schema checks.
• fast-decider-architecturesPuts a cheap fast model in front of a frontier model, with a gate that never serves a lower score.
• skill-forgeWrites or refines your own skills to a standard that holds up.