AI Guides › Workbench
By Nigel Guy · 6 min read
Most people ask Claude to "write this in a friendly, professional tone", accept the first draft, and wonder why it reads like everyone else's. Adjectives describe a voice; they do not contain one. The first draft also feels finished because it is tidy, which is exactly why the missing opinions and the borrowed phrasing slip through.
The rule: give Claude evidence of your voice and your own thinking, make it audit its draft against both, and never let it supply the opinions.
| Prompt | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| 1. Voice Profile | Turns your real writing into a reusable written description of how you write | Works on the free plan; Pro is $20 a month (about £15 at time of writing, check the £ price at checkout) | Setting up once per kind of writing | Only as good as the samples you paste |
| 2. Draft With Audit | Drafts from your notes, then checks the draft against the profile before showing it | Same | Emails, posts, updates, first drafts | It still cannot know facts you did not supply |
| 3. Claim Check | Marks every sentence as yours, borrowed, or invented | Same | Anything published under your name | Takes a minute longer; that minute is the point |
Prices and plan details move. Check claude.com/pricing before you rely on any figure here.
Paste three to five pieces you actually wrote and were happy with: emails you sent, posts you published. Avoid anything heavily edited by someone else or already generated by an AI. Different formats need different samples, so keep a separate profile for, say, client emails and for blog posts.
You are a writing analyst. Your job is to describe how I write so that a
future draft can be checked against it.
Context: below are [NUMBER] samples I wrote myself, all of type [KIND_OF_WRITING,
e.g. client emails]. My audience is [AUDIENCE].
[PASTE_SAMPLES_HERE, each separated by ---]
Steps:
1. Read all samples before you comment.
2. Describe my voice under these headings: sentence length and rhythm; how I
open and close; vocabulary I favour; words and phrases I avoid; how I
handle opinions, hedging and humour; punctuation habits; what I never do.
3. For each heading, quote a short real example from the samples. Do not
paraphrase and call it a quote.
4. List 5 "tells" that would make a draft sound NOT like me.
5. Finish with a profile of no more than 200 words that I can paste into
future prompts.
Constraints: describe only what is evidenced in the samples. If the samples
are too few or too similar to judge a heading, say so rather than guess. If I
have not told you the kind of writing or audience, ask before starting.
Self-check before answering: is every claim backed by a quoted example?
Remove any that are not.
Fill in: the number, the kind of writing, the audience, and the samples. Save the 200-word profile somewhere you can paste from.
Where to keep it: Claude has a Styles feature that can learn from uploaded writing samples, and Projects, where you can click "Set project instructions" and add files to the project knowledge. I confirmed the Projects steps in Anthropic's help centre. I could not load the official Styles page, so the exact Styles menu path is not stated here. Look for it in the style picker in the chat box, and treat a generated style as a starting point, not proof. A pasted profile in a Project works whichever way Styles behaves.
You are my drafting assistant. Write in my voice using the profile below.
Voice profile:
[PASTE_PROFILE]
Task: write [WHAT_TO_WRITE] for [AUDIENCE], about [LENGTH] long.
My raw material (the only source of facts and opinions you may use):
[PASTE_NOTES, DECISIONS, FACTS, MY_VIEW]
What a good result looks like: [ONE_SENTENCE_GOAL, e.g. the client agrees to
the new date].
Steps:
1. If anything in the material is missing that the draft needs (a date, a
name, my actual position), ask me. Do not fill gaps.
2. Write the draft.
3. Audit it silently against the profile: check the opening, sentence rhythm,
banned words, and each "tell" listed in the profile. Fix failures.
4. Show me the final draft, then a short list headed "Where I stretched": any
sentence that goes beyond my notes.
Constraints: no invented facts, figures, quotes or anecdotes. No phrases from
the "never" list. Output the draft first, the list second, nothing else.
Fill in: the profile, the task, your notes, and the goal. The "Where I stretched" list is the useful bit. It makes the model show its own overreach.
The trap is outsourcing the thinking. A model matches style well and supplies plausible-sounding opinions without being asked. A draft in your voice with someone else's reasoning is still not yours. Run this before publishing.
You are a sceptical editor. Below is a draft I plan to publish under my name,
followed by my original notes.
Draft:
[PASTE_DRAFT]
My notes:
[PASTE_NOTES]
Steps:
1. Split the draft into sentences.
2. Label each one: YOURS (traceable to my notes), BORROWED (a generic phrase
or stock argument), or INVENTED (a fact, number, example or opinion not in
my notes).
3. List every INVENTED item with the sentence it sits in.
4. For each BORROWED sentence, ask me one question whose answer would make it
specific to me.
Constraints: do not rewrite anything yet. If you cannot tell whether
something comes from my notes, label it INVENTED. Ask me for missing context
rather than guessing.
Self-check: did you label every sentence? Did you ask rather than assume?
Answer the questions in your own words, then ask for a revision that keeps your answers verbatim.