Review skill for your AI agent

Find the gaps that leave users stuck.

About to share a README, guide, skill, or site? Cold-eye gives your AI agent a repeatable review method. You get a report that quotes the problem, explains where a new user gets stuck, and tells you what to fix before the handoff.

Install Cold-eyeSee before, report, and after

You receive a Markdown critique with prioritized fixes. You decide which edits to make; reviewing leaves the original files unchanged.

A fictional notes app. Its setup guide omits a launch command that already exists in the app’s package.json.

1 · Before The user installs it, then gets stuck

Run npm install.
Launch the app.

2 · Cold-eye’s report The missing step, located and explained

Excerpt from README.cold-eye.md

F-001 · Missing launch command
README.md → Quick start

Cold reader: Installs the dependencies, then has to guess how to open the app.

Put: Add npm start, the command defined in package.json, and say how to recognize a successful launch.

3 · After you make the edit The guide now gets them to the app

Run npm install, then npm start.
Open the local URL printed in the terminal. You should see the Notes page.

A new user can finish setup without asking you for the missing step.

Illustrative review and author’s edit, not a recorded tool run. Sample commands belong to the fictional app.

Read the complete sample report →

What it reads, writes, and changes

Reads	A finished skill, spec, page, package, site, or repo
Writes	<name>.cold-eye.md next to a file, or cold-eye.md at a system root. Chat only, if you ask
Changes	Nothing, unless you separately ask for an edit

Verdict	Means
holds	A newcomer can run the job from the file
close	Almost. A few edits to the file would close the gap
fails a hostile read	The reader still has to invent too much

A repo, a package, or a site may get a split verdict: one for the main file read alone, one for all the files read together.

Clean verdict (illustrative)

Labeled example. A one-page checklist that names the file, the output path, and when to stop.

# Cold-eye — clean checklist

**Verdict:** holds

## Ranked changes

None.

## Protect

The four steps, and the line that says write `None.` when nothing is missing.

No invented faults. The file already tells a newcomer how to start, what to read, what to write, and when to stop.

The first-run sample

Get started has you run Cold-eye on a four-step sample checklist, until-ready.md. Step 4 says to repeat the review “until ready” and never says what ready means. A run should produce something with this shape:

# Cold-eye — review until ready

**Verdict:** fails a hostile read

## Ranked changes

1. **F-001** · test 10 — no_close
   Absent: the procedure, what “ready” means
   Cold reader: keeps repeating with no exit.
   Put: Name the readiness criterion, then hand the file over and stop.

How to check your result:

• The finding should point at step 4, the unresolved “until ready” condition.
• The checklist should be unchanged.
• Wording varies by model. Check the shape, not an exact match. A second finding, such as notes.md never being described, is fair if it points at the file.
• A corrected copy, until-ready-fixed.md, defines readiness and says when to stop. It should not get the step 4 finding.
• A failing verdict alone does not prove your agent loaded the skill. Check the agent's skill list, or ask it to quote the first heading of SKILL.md.

The repo's fixtures/ folder holds more sample subjects. Each has an expected.md naming the injected fault and the test that should catch it. The clean sample has no fault, so a wording-only finding on it is a miss.

What it checks

Ten questions, in order. Rank by what a newcomer hits first.

1. Can the reader run the job as steps: start, read, write, stop? Stages and principles are not a procedure.
2. Are required names present: schema, filenames, write-back, and what done looks like? A placeholder the procedure never binds is a guess.
3. Do the files match each other, and does what shipped match the page?
4. If a claimed check needs a second command, is that command next to the claim?
5. Does a definition include a worked example and a do-not-emit case?
6. Would the discovery copy fire on a different job?
7. Are the same rules copied until they drift?
8. Are maintainer notes sitting on a buyer page?
9. Does the procedure say write X while the named tool destroys X?
10. Does the procedure say when to stop?

Leave sentence polish. Cheerleading and unfinished plans are a different review.

Cold-eye is a readiness pass. It is not a code audit, a security audit, or a test run, unless those actions are the subject's own claimed checks. Detangler finds structural tangles after edits. Smell Check reviews prose register. Misemphasis reviews likely readings. A failed contract is not a preference about wording or layout.

Built by Catalyst Forge LLC. MIT.