Nothing to hand? — a docs page for a fictional fetch-scheduling library, seeded with the classic defects, plus the source file that exposes two factual errors.
Paste or drop the page
Paste the markdown, or drop the .md file straight onto the box — it is read in your browser and never uploaded. The instant lint flags style tells for free while you type — Title Case headings, hedging, trailing gerunds, marketing speak — before you run anything.
The AI grades it
Four rubric dimensions scored out of 10: Readability, Voice, Completeness, Accuracy. Every issue quotes the exact passage, names the rule it breaks, and gives the concrete edit. Paste the source code and accuracy is checked against the truth, not vibes.
Apply the fixes and go again
Tick the fixes you want and the app writes them back into your document, so you leave with a corrected .md file rather than a list of things to retype. Export the review with line numbers and suggestion blocks for your pull request, then review again and see which fixes landed.
Derived from the @tldraw/review-docs skill and the tldraw writing guides on SkillSafe.
What does Docs Critic actually score?
Four dimensions, each out of 10, for a total out of 40: Readability (sentence clarity, flow, paragraph length, scannable structure), Voice (hedging, passive constructions, Title Case headings, AI writing tells), Completeness (an opening definition, working examples for every concept, progressive disclosure, cross-references) and Accuracy (whether the technical content is correct). The total is computed in your browser from the four scores.
How can it check accuracy?
Only against something it can read. Paste the source code or API reference into the source material box and every API name, default value and described behavior in the doc is checked against it, with the mismatching line quoted. Leave it empty and accuracy findings are limited to what is provable from the page alone — self-contradictions and code that disagrees with the surrounding prose — and the review says so in its accuracy note.
Can it fix the document, not just describe the problems?
Yes, for the issues where a straight substitution is honest. Each issue that carries replacement text is shown with a checkbox; ticking it and choosing Copy corrected document or Download corrected .md splices those replacements back into your page at the exact character offsets the quotes were found at. Issues whose fix means adding a missing code example, deleting a section or restructuring have no checkbox, because no single substitution would be correct.
What is free?
The style lint is free and needs no sign-in: it runs entirely in your browser with no network call, and flags Title Case headings, hedging, trailing gerunds, formulaic transitions, marketing speak, AI vocabulary, negation parallelism, hollow importance claims, passive voice, wall-of-text paragraphs and unclosed code fences. Applying fixes, the Markdown export and the corrected document are free too. Only the graded review itself calls the model and costs credits.
Does my document get stored?
The page is sent to the model to be reviewed. If you are signed in, each review is also saved to your own history so re-reviews can show score deltas and which issues you resolved; those records are readable only by you. Dropped files are read in the browser and never uploaded anywhere.
What happens to a very long page?
Anything over 150,000 characters is clipped from the middle, keeping the opening and the ending, because a docs page defines its subject at the top and carries its options tables, cross-references and error handling at the bottom. The cut is announced to you above the review and to the model in the text itself, and the review is told not to report things missing that may have been in the removed span.