agentsclimarketplace

Research

Skill Alexanderdunlop/anvil-skills/skills/research

Turn the corrections anvil could not route into one research question aimed outside the repo, and write nothing at all.From its SKILL.md

Install
npx -y skills add Alexanderdunlop/anvil-skills --skill research

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

4.0 KB, as published. Nobody here has run it

/research

The outward channel. Every other skill improves this repo from this repo's evidence. This one turns what this repo could not resolve into a question for someone else.

This command writes nothing. No context file, no ticket, no log entry, not even an edit to unrouted.md. It emits a prompt to the session and stops.

That is not a small scope, it is the invariant. /improve is the only command that edits context files, and a second command reading the same evidence base and writing to disk would be a second writer under a different name. It must complete successfully with the whole of ${CLAUDE_PROJECT_DIR}/.anvil/ write-protected.

1. Precondition

If ${CLAUDE_PROJECT_DIR}/.anvil/ does not exist, stop. Tell the user this repo has no anvil setup and to run /setup.

Your inputs live under that directory too. Missing means there is nothing to research, not that you should go looking elsewhere.

2. Read exactly these

  • ${CLAUDE_PROJECT_DIR}/.anvil/feedback/unrouted.md — the primary input
  • ${CLAUDE_PROJECT_DIR}/.anvil/feedback/human.jsonl
  • ${CLAUDE_PROJECT_DIR}/.anvil/feedback/self.jsonl

No context file. Not CONFIG.md, not REVIEW_RULES.md, not any process/*.md. Your cost to the read budget is zero and it stays that way.

The parked pile is the question set. The logs are there so you can quote the full entry behind a bullet, and so you can see whether the same fingerprint has been parked more than once.

3. When it is worth running

Run it when the pile is past ten entries, or when the same fingerprint has been parked across several tickets. Those are the two shapes that mean something structural rather than something incidental.

Against an empty or near-empty pile, say so and stop. Do not produce a question anyway. A research prompt with no evidence behind it is the same error as seeding a context file from a guess: it looks like output and it is made up.

4. The prompt

One question. Not a list.

Pick the pattern with the most evidence behind it, and state:

  • the question, aimed at something outside this repo — how other projects route this kind of correction, whether a category is missing from the table, whether the thing being parked is a category error rather than a gap
  • the evidence, quoted by fb-NNNN id, with enough of each entry that a reader who has never seen this repo can judge it
  • what an answer would change — the routing table, a budget, a skill's instructions. If nothing would change, this is not the question worth asking

Aim outward. "Why does this repo keep parking verify entries" is a question /improve already has the evidence for. "Is 'how much of the suite to run' a process lesson or a config fact, and how do other projects decide" is a question this repo cannot answer from here.

5. Two destinations, and name both

  1. Run it yourself. Paste it into Claude research or a similar tool and act on the answer in this repo.
  2. Send it upstream. Open an issue on anvil-skills from the research issue template.

Upstream is the interesting one: it is the only path by which evidence from a repo anvil's author will never see reaches the routing table. It costs a GitHub issue and nothing else.

Nothing leaves this repo unless a human pastes it. There is no submission endpoint, no telemetry, and no network call in this command. Say so if the user asks where the prompt goes — the answer is "wherever you put it".

6. Then stop

Emit the prompt. Do not write it to a file, do not append it to unrouted.md, do not mark anything parked or dropped, and do not log the fact that you ran.

If acting on the answer produces a change, that change comes back through the normal route — a correction during a command, logged, promoted by /improve on evidence. Not from here.

Keep looking

Skills are one crate of 609,689. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.