Content Design Review / Claude Code skill
Review product copy.
Keep the final call.
A read-only review of product copy in your repository. Findings come first. You decide what changes.
Why I built this
I've reviewed product copy for over ten years: first as an editor, then as a content designer, and for five years teaching content design. Most of that time goes to the first pass: catching what's unclear, inconsistent, or simply untrue.
I wanted a second brain for that pass. Something fast that checks copy against what the product actually does, separates confirmed problems from suggestions, and leaves the final call to me.
The rest stays with me. Quality, taste and judgment are where a content designer makes the difference, so the final call is never automated.
A finding, not a silent rewrite
ONB-3, ONB-1 · index.html:20
(onboarding step 2)
Original
Log entries with species, date, location and notes. Sync your records across all your devices automatically. Share observations with friends and the Field Log community. Export your data whenever you want. Get smart species suggestions powered by AI.
Issue
The README and code show no server, no account and no sync, and there's no code for sharing or suggestions.
Rewrite
Log each sighting with species, date, location and notes. Sightings are saved in this browser, on this device only. You can export them as a CSV file at any time.
Files changed: none
Field Log is a small, deliberately flawed demo app in the repository. This finding checks an onboarding promise against what its code and README actually support.
Review first. Approve by ID.
The skill reads UI source, project docs and any voice guide or glossary. It traces what the code does before suggesting copy.
Hard checks identify usability failures you can show from the source. They become confirmed problems. Pref checks cover context-dependent choices. They can only become suggestions, with the assumed audience and a concrete cost to the person stated.
Each finding includes an ID, severity, rule, file and line, exact original text, issue and proposed rewrite. Hard or Pref is the check type; major is a separate severity rating.
Approve specific rewrites by ID, for example apply F1, F3 and F4. The review itself does not create, edit, move or delete files. Until you ask for changes: Files changed: none.
Install in Claude Code
From the root of the project you want to review, run these two commands from the README:
git clone https://github.com/marcustrell-cat/content-design-review.git /tmp/content-design-review
mkdir -p .claude/skills && cp -R /tmp/content-design-review/skill/ux-copy-review .claude/skills/content-design-reviewThen type /content-design-review in Claude Code. If the skill is not listed when you type /, restart the session.
What v0.2 changes
v0.2 adds a filter: if the skill’s own reasoning says the copy works in context, it doesn’t report the finding. It still always reports clashes with the product’s terminology and error messages that people actually see.
7.8% flagged
26 of 39 findings came with the skill’s own admission that the case was weak (‘works in context’, ‘cost is small’).
4.6% flagged on average
Range 3.6% to 5.6%. The strongest catches stayed, like an unexplained ‘must NOT have additional properties’ error and a vague ‘Yes, I understand’ on a dangerous action.
The drop is not an accuracy score: the filter removes findings the skill itself considers weak. A fine-tuned model flagged 9.9% on a similar sample.
The results, with limits
The skill has been tested on three small apps kept in the repository (one with good copy, one with a warm voice, one with deliberately bad copy) and on a sample of real PostHog copy. The results:
| Test | Runs | Result |
|---|---|---|
| Good copy | 8 | All passed. Six later runs had zero findings. |
| Warm voice | 6 | All passed. One later run still made a minor overreach. |
| Bad copy | 10 | 18/18 items it had to catch, in each completed run. |
| PostHog UI strings (500 random, real product) | 3 | flags 4.6% on average (range 3.6–5.6%). A fine-tuned model flagged 9.9% on a similar sample. |
On good copy, a run passes when the skill raises no major findings and at most two minor ones. Smaller notes are still logged.
Early bad-copy runs cited wrong lines, and one stated an error cause it had not verified. One run was stopped because the examples it was given reused text from the test copy. The five completed runs that followed showed none of the failures the tests rule out, including invented facts or file changes.
The first two good-copy runs found defects in the test copy itself, which were fixed. A warm-voice run made two taste-based suggestions, which prompted a stricter rule for suggestions.
The first three rows come from those older test apps (draft and v0.1, scored by hand, not re-run for this page); their scoring details are in the bad-copy log, good-copy log and warm-voice log, without the full review transcripts or file diffs. The PostHog row is new: three runs on v0.2. Treat all of these numbers as recorded evidence, not proof of broad reliability.
English only. The checks are written for English UI text.
Repository text only. It reviews copy that lives in the repository: in the code and in translation files. Text the server generates at runtime, text inside images and emails absent from the repository are not covered.
Visual design, contrast, layout, code quality, legal review and translation quality are outside scope.
About a quarter of findings change between runs on the same strings. Read the rate as a range, not a score.
The PostHog sample follows the original author’s method, but it isn’t the exact same set.
Keep the final call.
Read the skill and its evidence before trying it on your own copy.
View on GitHub