setup-researchlisted

# Setup Research Project Record this project's conventions into its root `CLAUDE.md`, so `handoff` and `design` stop guessing at them. `CLAUDE.md` is auto-loaded every session, so nothing has to remember to read a config file. 1. **Dig before asking.** Read `README`, `pyproject.toml` / `requirements.txt` / `environment.yml`, `Makefile` / `justfile` / `scripts/`, `configs/`, `.gitignore`, `git log --oneline -30`. Deps reveal the training stack; `scripts/` usually holds the real launch commands. 2. **Ask only what's left** — one question at a time, each with your recommended answer. 3. **Never invent a command.** Confirm the script or entry point exists; if you can't, record it tagged `unverified`. 4. **Write the block** into the root `CLAUDE.md`, wrapped in `<!-- research-conventions:start -->` / `<!-- research-conventions:end -->` so re-running updates only this block. Drop any subsection with no real answer — an honest gap beats a plausible fabrication. ```markdown ## Research project conventions ### Docs - `handoff.md` — current state, overwritten (path) - `CHANGELOG.md` — experiment history, append-only (path) - `Design.md` — current pipeline design (path) ### Environment - Activate: <conda / venv / uv> - GPUs: <count, launcher and its flags> ### Commands - Train / Eval / Quick debug run — full command lines ### Layout - Configs / checkpoints / logs / datasets ``` 5. **Version control.** `CHANGELOG.md` and `Design.md` are the project's memory and belong in git; ask