A Claude Skill for grading writing on five dimensions, with a focus on catching AI-shaped prose in 2026.
Runs on a document you wrote, a document you received, or a document you're about to ship. Produces an HTML critique with letter grades, two percentages, a redline, and a version-history table.
Five dimensions. Three letter grades. Two percentages.
- Professionalism (A through F). Fit for the venue. Tone, register, completeness.
- Writing Style (A through F). Rhythm, clarity, sentence variety, voice consistency.
- Leadership Capability (A through F). Position-taking, ownership, action for the reader.
- AI Authorship (0 to 100 percent). How likely a thoughtful reader is to flag this as AI-generated. Lower is better.
- Snarky Meter (0 to 100 percent). Is the spice calibrated for the audience.
The first three are quality. The last two are pattern detection. Together they tell you whether a piece is fit to ship, fit to send, fit to forward, or fit for the trash.
This is not an AI detector in the forensic sense. It's a writer's rubric that includes an AI-authorship dimension. The score is a reader's judgment, not a forensic verdict. Don't use it to accuse anyone of anything. Use it to evaluate writing on its own terms.
This is not a substitute for a voice book. The rubric runs against rules. If you don't have rules (refused words, refused openers, refused sentence shapes, refused closers), the rubric grades against generic AI tells. That's useful, but it's the generic version. The full value comes from running the rubric against your own voice rules.
Two paths, depending on which AI you're using.
# From your claude-skills directory
cd ~/Code/claude-skills/skills
git clone https://github.com/brianbeals/voice-critique.gitRestart Claude Code, then invoke:
/voice-critique [path to document]
The skill gathers context, then produces an HTML output you can open in a browser, print, or attach to an email.
Open prompt.md in this repo. Copy everything between the rules near the top. Paste it into your AI chat as a single message. Then paste or attach your document.
The AI will ask you a few context questions, then return a complete, self-contained HTML critique in a single code block. Same five-dimension rubric, same grading scale.
Where you see it rendered depends on the tool. Claude.ai renders the HTML inline as an artifact. Gemini, ChatGPT, and Copilot show the HTML as source instead of rendering it; copy the code block, save it as critique.html, and open it in any browser. For inline rendering with no save step, use Claude.ai or the Claude Code skill (Path A), which writes the file straight to disk.
If you just want the prompt without opening this repo, there's a copy-and-go page at brianbeals.com/tools/voice-critique. One button copies the same prompt; paste it into any AI chat.
The Claude Code path is more powerful (file I/O, version-history support, optional voice-rules.md). The chat path is more accessible (works in any AI tool you already have open).
/voice-critique ~/Documents/draft-essay.md
Use the critique to decide what to edit before shipping. Run it again after edits to see the trajectory.
/voice-critique ~/Downloads/vendor-pitch.docx
The AI Authorship percentage flags writing patterns that may make the pitch feel AI-shaped. The other dimensions tell you whether the pitch is worth a meeting.
Run it after each edit. The skill produces version-numbered output. The trajectory across V1, V2, V3 is the story.
A self-contained HTML file with:
- A grade summary card at the top
- One section per dimension with the letter grade or percentage, a gauge, and the reasoning
- A version history table (if there are prior versions)
- A redline section with specific original-and-suggested pairs
- A final verdict line and a ship-readiness flag
Open in a browser to read interactively. Print clean for a paper copy. Attach to email for sharing.
- Edit your document. It produces a critique, not a rewrite.
- Grade content accuracy. The rubric evaluates writing, not facts.
- Replace your voice rules. The rubric needs rules to grade against. The default is generic; the value is in customizing.
The default rubric grades against generic AI tells and broad professional norms. To get more value, write your own voice rules and point the skill at them.
See docs/voice-rules-example.md for a one-page voice book template. Replace the generic patterns with your refused words, refused openers, refused sentence shapes, and refused closers. The skill reads from that file when grading.
This skill came out of writing a series of essays about voice discipline in the AI era. The essays are at https://brianbeals.com/writing. The relevant ones:
- "The Last Moat Is Your Voice" (the strategic argument)
- "AI and the Poisoning of Professionalism" (the failure mode at the email level)
- "An LLM With a Logo" (the failure mode at the brand level)
- "How I Review My Own Writing" (the methodology this skill implements)
If you read the essays first, the skill makes more sense.
MIT. Use it, fork it, adapt it, sell consulting around it. The framework is generic by design. The value is in the calibration to a specific voice, which is the work this skill cannot do for you.
If you want a version of this calibrated for your company's voice (refused words, refused shapes, audience-specific snark levels, dimension weighting tuned to your content velocity), that's work I do for enterprises. brian@brianbeals.com.
Brian Beals. brianbeals.com. The rubric reflects how one writer reviews his own work and the writing that lands in his inbox. The implementation is open. The voice rules behind any specific calibration are not.
© Brian Beals · brianbeals.com