TL;DR
Most “AI writing” tutorials show a one-liner prompt and a polished result, hiding the part where the first draft gets overwritten three times and still sounds generic. The fix is a five-part prompt — audience, thesis, evidence, voice sample, format — plus a two-pass revision loop and Claude’s Styles feature to lock your tone. This gets you shippable prose on the first or second draft for blog posts, essays, scripts, and short reports. Below is the exact structure, the on-screen settings to use, and the checklist that strips the AI tell before you publish.
What you need
This works on the free tier, but the paid plan matters for length and consistency. As of June 2026:
| Plan | Price/mo | Models | Why it matters for writing |
|---|---|---|---|
| Free | $0 | Limited Sonnet 4.6 | Fine for one short piece a day; tight message caps |
| Pro | $20 ($17 annual) | Opus 4.7 + Sonnet 4.6, Projects, Styles | The sweet spot for regular writers |
| Max | $100 / $200 | Same models, 5x / 20x usage | For high-volume or long-document work |
Pro at $20/month unlocks Claude Opus 4.7 (the more direct, opinionated model that follows instructions literally) and Claude Sonnet 4.6 (the faster workhorse), plus Projects and the Styles feature this workflow leans on. Both models carry a 1M-token context window, and Opus 4.7 can write up to 128K tokens (~96,000 words) in a single response — so a full long-form draft fits in one turn. See the Claude beginner guide for account setup.
Use Opus 4.7 for the draft (its prose is sharper and it hedges less than Sonnet) and Sonnet 4.6 for fast line edits where speed beats polish.
Before you start
Gather four inputs. Skipping any one is the most common reason drafts come out generic:
- Audience, precisely. “Developers” is too vague. “Senior backend engineers evaluating a new database” is workable.
- Thesis, one sentence. If you cannot write it in a sentence, you do not have a thesis yet, and Claude cannot invent one for you.
- Five evidence points: data, quotes, anecdotes, examples. Without these, the draft is opinion-flavored air.
- A voice sample: 2-3 paragraphs of your prior writing, or an author you want to sound like. This is the single highest-leverage input and the one most people skip.
The five-part prompt
Claude is trained on XML-tagged prompts, so wrapping each section in tags gives cleaner, more organized output than one long paragraph. Opus 4.7 follows instructions more literally than older models, which makes clear boundaries matter even more — without tags it may treat your evidence as instructions or your example as a request.
- State audience, tone, and goal in one block:
<audience>
Senior product managers at SaaS companies, 5+ years experience,
skeptical of AI hype, decision-makers on tool purchases.
</audience>
<tone>
Confident, plain English, dry humor allowed, no marketing adjectives.
</tone>
<goal>
Convince readers that AI prompt libraries are worth building internally,
not buying from a vendor.
</goal>- Give the thesis as one tight sentence inside its own tag:
<thesis>Internal prompt libraries beat vendor tools because they encode tacit team knowledge that vendors literally cannot ship.</thesis> - Provide five evidence bullets in an
<evidence>block. Concrete numbers, named examples, specific anecdotes. No generalities. - Paste a
<voice_sample>— 2-3 paragraphs of writing you want the draft to mirror. - Specify format explicitly: word count, paragraph count, intro/body/close structure, and any required elements (a subhead every ~200 words, etc.). Vague length requests are why drafts run long and shapeless.
Lock your tone with Styles
The Styles feature (in the compose toolbar, on Free and Pro) is the part most people miss. Built-in presets — Normal, Concise, Explanatory, Formal, Learning — tune tone and detail level per chat. More useful: create a custom style from a writing example. Paste a few hundred words of your own writing and Claude derives a reusable style you can apply to every draft, instead of re-pasting the voice sample each time.
Note for June 2026: Anthropic is migrating Styles into Skills. Custom styles you create are carried over automatically (disabled by default — re-enable under Customize > Skills) and are invoked as slash commands like /your-style-style. The Concise, Explanatory, and Formal defaults are being retired; Learning survives as a skill. If your Styles menu has already changed, look under Skills instead.
Layer three controls and they stack on every reply: account-wide Instructions for Claude, a per-chat Style, and per-workspace Project instructions if you write inside a Claude Project.
The two-pass revision loop
Revise in two separate passes. Mixing them makes you rewrite the same paragraph four times.
- Pass 1 — structure. “Move the counterargument earlier.” “Cut the third example, it repeats the first.” Fix the skeleton before any sentence-level work.
- Pass 2 — line edits. “Kill three adverbs in paragraph two.” “Vary the sentence length in the intro.” Only after structure is locked.
The first-run exercise
The fastest way to calibrate your inputs:
- Pick a piece you have already written and shipped. Reverse-engineer the prompt: write the audience, thesis, evidence, voice, and format that would have produced it.
- Run that prompt with Claude. Compare its draft to your real piece.
- Name the gap — voice, structure, evidence weighting, or all three. The gap tells you which input to strengthen.
- For your next real piece, lead with whichever input was weakest in the test.
The pre-publish checklist
This is where the AI tell gets caught. Run every item:
- Read it aloud. AI cadence shows up as uniform sentence length and missing rhythm. Fix as you read.
- Verify every statistic, name, and quote. Claude will confidently invent the occasional source — and hallucination rates are non-zero even on Opus 4.7. Treat any specific number, date, or citation as a placeholder until you confirm it.
- Strip the AI tells: triplets (“efficient, scalable, and powerful”), hedge phrases (“it’s important to note”), meta-commentary (“let’s dive in”), and em-dash overuse.
- Check the thesis is still on the page. AI drafts drift toward balanced both-sides framing that dilutes your point.
Reusing the workflow
- Save each publication’s audience + tone + voice blocks as a Project or a custom Style. Swap thesis and evidence each time.
- Build a small voice library — 3-5 styles for different contexts (newsletter, blog, internal memo, deck). Apply the closest match per piece.
- Keep a short list of phrases Claude overuses in your work and paste it into Instructions for Claude.
- Re-run a test piece each quarter. Model voice shifts subtly with each release; Opus 4.7’s default tone is noticeably more direct than Opus 4.6’s.
Common mistakes
- No thesis. Without a spine the draft can be technically “good writing” and still say nothing.
- Asking for “good writing” without naming audience, tone, or length. Claude defaults to generic blog voice.
- Skipping the voice sample or a Style. The highest-leverage input is the one most people omit.
- Letting Claude invent evidence. Anything specific it generates without a source is a placeholder until verified.
- Mixing structure and line edits in one pass.
- Publishing without reading aloud. AI cadence is audible; reading aloud catches it.
FAQ
- Will my readers know it was AI-written?: They will if you skip the voice sample and revision passes. With a custom Style plus two-pass editing, the gap closes substantially. The read-aloud check is what separates “obviously AI” from “human.”
- How long should the prompt be?: For a 1,000-word piece, expect a 400-700-word prompt. The voice sample is most of that length — or zero, if you have saved it as a Style.
- Opus 4.7 or Sonnet 4.6 for drafting?: Opus 4.7. Its prose is sharper and it hedges less. Use Sonnet 4.6 for quick line edits where speed matters more than polish.
- Can Claude write in my exact voice?: Closer than you expect with a strong voice sample or custom Style. Not identical without a round or two of iteration.
- What about hallucinated statistics?: Common, and not fixed by a better model. Treat every specific number, date, or quote as a placeholder until verified.
- Do I need Pro, or is Free enough?: Free works for an occasional short piece. Pro at $20/month is worth it once you write regularly: you get Opus 4.7, Projects, higher limits, and saved Styles.