agentleFS
ArchAstroArchAstro/archdev/eval-review/SKILL.md

eval-review

Skill4 starsChanged 41 days agoMITView on GitHub

Use when a person needs to judge a set of outputs one at a time in the browser — rating model or agent outputs against criteria, comparing two candidates blind, labelling examples, or spot-checking generated summaries, diagrams or code changes. You write one JSON file with the items and the questions, host it locally with the bundled script, and ArchDev shows a keyboard-driven review deck with each item's picture, description, pull request and code; ratings save to a file you read back. Triggers include "have a human rate these", "set up an eval review", "let me go through the examples", "which of these is better", "human eval", "label these", and "collect feedback on these outputs".

2 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions2 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 2

    1. buildomator/buildomator★ 102same name
    2. alexaiarchitect/google-adk-skills★ 0same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository6 other files it gives its agents.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 2 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Use it when a person's judgment is the measurement: rating outputs, choosing between candidates, labelling a sample, or checking a claim …

    line 13
  2. Very specific:

    Follow the archdev skill for activity reporting.

    line 20
  3. Distinctive:

    Ask what the person can judge from the card.

    line 73
  4. Very specific:

    Add a follow prompt on the answers that mean something is wrong.

    line 77
  5. Very specific:

    Put per-item facts in hints (word counts, a classifier's label), not in the criterion title.

    line 81
  6. Very specific:

    If the server warns that the site did not serve /eval-frame, that ArchDev deployment does not have eval review yet.

    line 114
  7. Distinctive:

    When the person says they are done, or partway:

    line 141
  8. Very specific:

    Run --summary for the counts, then read the ratings file for the why text, flags and notes.

    line 143

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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

Eval review

Turn "can you look through these and tell me which are good" into a review a person finishes quickly. You prepare the items and the questions; the person rates them in ArchDev, one card at a time, with number keys. Their answers, flags and notes land in a JSON file beside yours.

Use it when a person's judgment is the measurement: rating outputs, choosing between candidates, labelling a sample, or checking a claim against code. Do not use it to review a diff for merge (archdev inspect local) or a task plan (the tasks skill).

Activity and risk ownership

Follow the archdev skill for activity reporting. Plan/task lifecycle events retain their risk.plan / risk.task assessments. Server automation owns hosted PR and code-region grading: report PR events without fetching, waiting for, reusing, or attaching a server grade. No local assessment, annotation or stop hold is required.

What the reviewer gets

  • One item per card, a progress strip, and the next few items listed.
  • Your criteria as rows of options. 1-9 answers the highlighted row, ↑ ↓ moves between rows, J/K moves between items, ↵ goes on once every row is answered.
  • An answer can ask for one line of explanation, and every item has a notes box.
  • The item's picture, with Z to enlarge. Parts of it can be flagged.
  • References: a pull request or a file opens in a side panel with the description and the diff; D opens all the item's code.
  • A summary at the end with tallies and whatever is still unrated.

1. Write the eval file

One JSON file holds everything. The full field list, the three exhibit kinds and a complete example are in references/format.md; read it before writing your first file. The shape:

{
  "title": "Release note drafts",
  "criteria": [
    {
      "key": "accurate",
      "title": "Does it match the change?",
      "options": [
        { "value": "yes", "label": "Yes" },
        { "value": "partly", "label": "Partly" },
        { "value": "no", "label": "No" }
      ],
      "follow": { "no": "What is wrong?" }
    }
  ],
  "items": [
    {
      "key": "pr-482",
      "title": "Checkout retries a declined card once",
      "exhibit": { "kind": "image", "mediaType": "image/svg+xml", "data": "<svg …>", "alt": "Draft note" },
      "description": "Markdown shown beside the card.",
      "files": [{ "filename": "src/charge.ts", "status": "modified", "additions": 12, "deletions": 3, "patch": "@@ -1,3 +1,4 @@\n …" }]
    }
  ]
}

Writing good criteria:

  1. Ask what the person can judge from the card. "Would you understand this from the picture alone?" works; "Is this good?" does not.
  2. Two to nine options each, in the order a reader expects. A 1-5 scale is five options; label the ends ("1 wrong", "5 exact").
  3. Add a follow prompt on the answers that mean something is wrong. That one line is usually the most useful thing you get back.
  4. Three or four criteria is comfortable. More than seven slows the reviewer enough that they stop.
  5. Put per-item facts in hints (word counts, a classifier's label), not in the criterion title.

Choosing the exhibit:

The person is judgingUse
One picture, card or screenshotimage (SVG text or base64 PNG)
Which of two to four candidates is bettercompare; shuffle the order per item yourself and keep the key in a separate file
Something they need to click inside, or a layout only you can drawframe: your own HTML and script in a sandbox, see references/format.md
Text onlyframe with the text as HTML, or no exhibit and the text in description

For anything about a code change, include the pull request as a reference and the diff as files, so the reviewer can check a claim without leaving the card. Split a unified diff per file: each file's patch starts at its first @@ line.

2. Check it, then host it

Resolve the absolute directory containing this loaded SKILL.md; the script is beside it. It needs Node 18 or newer and nothing else.

node <skill-dir>/scripts/serve.mjs eval.json --check   # lists every problem in the file
node <skill-dir>/scripts/serve.mjs eval.json           # hosts it and opens the browser

Fix everything --check reports before hosting; the web app rejects a session it cannot show. Run the server in the background and leave it running: it is what the browser reads from and saves to. It listens on loopback only, and the link it prints carries a one-time token for this session. Do not paste the link into a pull request, an issue or a chat channel.

If the server warns that the site did not serve /eval-frame, that ArchDev deployment does not have eval review yet. Tell the person; do not work around it.

Tell the person what you are asking them to judge, how many items there are, and that their answers save as they go. Then wait. Do not rate items yourself.

3. Read the ratings

Ratings are written to eval.json.ratings.json after every change:

{
  "pr-482": {
    "answers": { "accurate": "no" },
    "why": { "accurate": "says it retries twice" },
    "flagged": ["retry-count"],
    "note": "",
    "at": "2026-11-12T15:44:09.194Z"
  }
}
node <skill-dir>/scripts/serve.mjs eval.json --summary   # tallies and what is still unrated

When the person says they are done, or partway:

  1. Run --summary for the counts, then read the ratings file for the why text, flags and notes. The explanations matter more than the tallies.
  2. Report what they found in their words, grouped by what it means for the work: which items failed and why, what was flagged most, what they asked for.
  3. For a blind comparison, join the answers with your key file before reporting; never show the key to the reviewer first.
  4. Say how many items are unrated. A partial pass is a result, with its size stated.

Restarting the server keeps the saved ratings and makes a new link; a tab opened from the old link cannot save, so send the new one.

Download now (ZIP)

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.