Use when a problem is genuinely HARD and needs more reasoning power than routine work — subtle correctness bugs, tricky algorithms, gnarly multi-file diagnosis, non-obvious async/concurrency issues, or architecture/design tradeoffs. It runs on Fable (a more capable tier than the Opus main session), so delegate the…
Use BEFORE writing code against any external service, API, library, or tool the team hasn't already documented — Anthropic/Claude SDK, MCP, LangGraph, GEO/PubMed, deploy platforms, a new dependency, etc. Reads the OFFICIAL documentation, extracts the exact signatures/params/version constraints, and records them in the…
Use AFTER any code change to confirm nothing broke — before calling work "done" and before any commit/push. Runs the test suite plus a golden-mode smoke run fully OFFLINE (no API key, no cost, deterministic) and reports pass/fail honestly. It cannot edit code or tests by design, so it can never "fix" a failure into a…
Use BEFORE every git push (and any time you're unsure what's about to be committed) on this PUBLIC repo. Scans staged and tracked content for secrets, credentials, and personal/clinical data, flags anything risky, and asks the user before proceeding. It reports and blocks — it never edits files or pushes on its own.
Put TransBench into golden mode and hand the user a fresh single-use Claude Science link to record the demo against. Invoke when the user says "link to record skill", "record link", "golden link", "give me the recording link", or "/record-link". Golden mode makes generateexperiment return the pre-captured lupus brief…
Claude Code instructions for kayomarz97/TransBench, covering claude.md — working agreement for transbench, project at a glance (the what), commands (exact — don't guess), rules (non-negotiable) and the agents (and exactly when to use them).