agentleFS
marin-communitymarin-community/marin/.agents/skills/research-engineering/SKILL.md

research-engineering

Skill4.1k starsChanged 49 days agoApache-2.0View on GitHub

Explore Marin designs, datasets, and experimental pipelines through iterative analysis and experiments; preserve a runnable research branch and notebook, then promote selected code into Marin PRs. Use for experimental programs, not bounded background research or routine debugging.

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 3 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill when answering a question requires exploring alternatives, analyzing data or artifacts, or iterating on experiments before …

    line 8
  2. Distinctive:

    Read the research engineering guide for the record, branch, and promotion workflow.

    line 13
  3. Very specific:

    Start from the decision, baseline, constraints, and useful stopping condition.

    line 21
  4. Very specific:

    Search prior work through use-echo and inspect the current checkout.

    line 22
  5. Very specific:

    Use a task issue in marin-community/marin-logbooks as the public entry point.

    line 23
  6. Distinctive:

    Keep the issue body current and its comments useful as a readable account of the important attempts, results, and changes of direction.

    line 33
  7. Very specific:

    Use the domain skill for each concrete operation, such as add-dataset, write-pipeline, launch-rl, add-pallas-kernel, or profile-training.

    line 41

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Research engineering

Use this skill when answering a question requires exploring alternatives, analyzing data or artifacts, or iterating on experiments before choosing an implementation. Work can start before any run is launched. Background research alone follows experiment records.

Read the research engineering guide for the record, branch, and promotion workflow. Read the current marin-logbooks guidelines before creating or updating its branches; research branches inherit Marin's root instructions rather than the logbooks index branch's instructions.

Coordinate the work

  • Start from the decision, baseline, constraints, and useful stopping condition. Search prior work through use-echo and inspect the current checkout.
  • Use a task issue in marin-community/marin-logbooks as the public entry point. Reuse an existing issue; when the task authorizes a new public research program, use file-issue with target repository marin-community/marin-logbooks to create its coordinating issue. An existing Marin issue can link to it. Do not publish private work without authorization.
  • Work on a Marin-derived branch in logbooks. Preserve the notebook and supporting analysis alongside the actual experimental source changes.
  • Record why each meaningful alternative was tried, the evidence it produced, what was learned, and the next decision. Commit and push supporting code and notebook entries before a milestone issue update or handoff.
  • Keep the issue body current and its comments useful as a readable account of the important attempts, results, and changes of direction.
  • Extract selected code into Marin contribution branches. Use manage-pr for their validation, publication, and feedback; keep the research branch intact. A useful negative result can finish a program without a code PR.

Delegate execution

Use the domain skill for each concrete operation, such as add-dataset, write-pipeline, launch-rl, add-pallas-kernel, or profile-training. use-iris owns job operations and query-finelog owns telemetry queries. This coordinating workflow does not authorize compute launches or recovery actions beyond the task's scope. Record meaningful restarts under their actual execution identities.

Babysit launched runs and PRs through a heartbeat: a scheduled follow-up in the host application that resumes monitoring after the active turn. Keep one monitor per task with its execution or PR identity, latest observation, owner, and exit condition. Report meaningful changes and required action; stay quiet while unchanged. Use use-iris for job completion checks and manage-pr for PR feedback and exit conditions.

Download now (ZIP)

Trending

More talk lately.about the whole repository

Dec 2025: 1 mentionJan 2026: 0 mentionsFeb 2026: 1 mentionMar 2026: 1 mentionApr 2026: 1 mentionMay 2026: 0 mentionsJun 2026: 2 mentionsJul 2026: 1 mentionAug 2026: 6 mentionsSep 2026: 13 mentionsOct 2026: 25 mentionsNov 2026: 0 mentions
Dec 2025Jun 2026Nov 2026
Mentions a month of the marin-community/marin repository that our weekly crawl found. There is no adoption data for this file yet.

Who's talking about marin-community/marin

The biggest 5 of 55 mentions linking to the repository that our crawl found. Each line is their own headline.

See all 55

Talked about on GitHub and Bluesky (55 times).

On GitHub · 50

45 more on GitHub

On Bluesky · 5

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.