Reusable operating instructions for coding agents in filesystem-first software and research projects.
This is not an app or framework, but rather, a compact policy bundle for agents that work inside real project directories, inspect local files, respect dirty worktrees, run commands, track jobs, and leave durable evidence for the next session.
AGENTS.md- global operating contract for developer agents.RESEARCH.md- research, eval, and paper discipline.RUNS.md- long-running jobs, GPU use, and run provenance.TOPICS.md/topic-definitions.md- topic vocabulary and curated jargon.skills/- reusable workflows for handoff, review, task start, and shipping.
Private machine-local preferences live outside public committed instructions;
in this setup, that means AGENTS.user.md.
Do you want a strict procedure that scripts today's weaker agents, or lean policy that preserves real intent for strong present and future agents? This repo chooses the latter.
It tells agents to:
- preserve continuity and resumability across agent sessions;
- verify project claims against the repo;
- preserve unrelated WIP in dirty trees;
- route global, project, and branch rules to durable files;
- write commits that preserve motivation and decisions;
- use task files and topic docs as working memory;
- record provenance for research and long-running runs.
It can also stand alone as a repo-local AGENTS.md or CLAUDE.md.
- Developer: inspect first, edit narrowly, verify, commit cleanly.
- Researcher: maintain experiment logs, eval splits, artifact provenance, and paper-facing claims for intellectually honest, reproducible public results.
- Run operator: launch, monitor, and summarize agent-driven jobs so long-running work survives handoff.
The common assumption is a real filesystem working tree: source, tasks, logs, artifacts, and git history are available for agents to inspect.