Claude as a writing assistant

How to get genuinely good writing out of Claude - style control, Projects for voice consistency, editing workflows, and five reusable prompt templates for real writing jobs.

Claude writes well when you give it a brief rather than a topic. What separates a usable draft from generic filler is almost never the model — it is whether you supplied an audience, a purpose, a length, a voice sample and a list of things not to say. This page covers how to set that up, the editing loop that beats asking for a perfect first draft, and five templates you can copy.

It also covers where Claude is weak, because that determines how much editing to budget for. Checked against Anthropic's documentation on 21 August 2026.

#Give it a brief, not a topic

“Write a blog post about remote onboarding” has no constraints, so the model fills the gap with the statistical average of everything it has read on the subject — the flat, listicle-shaped prose people complain about. A brief removes the ambiguity: who is reading, what they should do afterwards, what they know already, and what has been said too often.

The most useful line in any writing prompt is the negative one. Telling Claude what not to include changes the output more reliably than any amount of encouragement.

#Template 1 — long-form article brief

Write a 1,400-word article. Follow this brief exactly.

READER: [who they are, what they already know, what they suspect]
PURPOSE: after reading, they should [specific action or belief change]
ANGLE: the one claim this piece makes: [claim in a single sentence]
EVIDENCE I AM GIVING YOU: [facts, numbers, quotes - with sources]
VOICE: [2-3 adjectives] - see the sample below
STRUCTURE: open with [X]; the turn happens at [Y]; end on [Z]
LENGTH: 1,400 words, sections of no more than 250 words

DO NOT:
  - open with a rhetorical question or a definition
  - use "in today's fast-paced world", "delve", "landscape", "unlock"
  - invent statistics, studies, quotes or examples
  - cover [topics already handled elsewhere]

If a claim in my brief needs evidence I have not given you, mark it
[NEEDS SOURCE] instead of filling the gap.

VOICE SAMPLE:
"""
[300-500 words of your own writing]
"""

The [NEEDS SOURCE] rule is what separates this from an ordinary prompt: unsupported claims come back flagged rather than written confidently for you to catch later.

#Use Projects to hold voice, style guides and reference material

Retyping a brief every session is how voice drifts. A Project in claude.ai is a workspace with its own knowledge base and custom instructions, so the style guide, the terminology list, three strong past pieces and any source material stay in context for every chat inside it. Custom instructions carry the standing rules — spelling convention, sentence length, how to handle first person. You then write the brief for the piece, not the whole apparatus.

#Use Artifacts to iterate on a document, not regenerate it

Substantial, self-contained output opens in an Artifact — a separate panel holding the document instead of burying it in the chat scroll. For writing, that is the difference between editing and re-rolling: “tighten section three and leave everything else alone” returns the same document with one section changed, not a fresh draft that has quietly lost the two paragraphs you liked.

Change one thing per turn so regressions are traceable, copy out any version you are happy with, and keep structural edits separate from line edits. Artifacts need code execution and file creation enabled in settings.

#Draft, critique, revise — the loop that beats one perfect prompt

Asking for an excellent first draft rarely produces one. Three cheaper turns produce more: get a draft, ask for a critique in a separate prompt with no permission to rewrite, then hand that critique back for a revision. A model asked to critique and fix at once tends to reword lightly and declare victory. Running the critique in a fresh conversation helps further — a clean context has no investment in the draft.

#Template 2 — editing and critique pass

You are a demanding editor. Critique the draft below.
Do not rewrite it. Do not soften anything.

Report, in this order:
1. Does it deliver on its stated purpose? One sentence, yes or no.
2. The three weakest paragraphs, quoted, with what is wrong with each.
3. Every claim that needs a source or is unsupported.
4. Anything a hostile, well-informed reader would attack first.
5. Where the argument actually turns - and whether that is too late.
6. Sentences that could be cut with no loss. Quote them.
7. Any tic: repeated sentence shapes, triads, hedging, "not just X
   but Y" constructions, throat-clearing openers.

Skip anything that works. I want problems, not a balanced review.

PURPOSE OF THE PIECE: [restate it]
DRAFT:
"""
[paste]
"""

Then, next turn: “Revise against points 2, 3 and 6 only. Leave the structure alone. Show the revision, then list what you changed.” Naming the points stops a critique becoming a rewrite.

#Matching an existing voice

Claude imitates a voice it can see far better than one you describe. “Conversational but authoritative” means almost nothing operationally; 500 words of the actual writing means a great deal. Make the analysis explicit first, so you can correct a wrong reading before it is baked into a draft.

#Template 3 — style matching

Below are three samples of writing I want you to match.

First, analyse them. Tell me:
  - average sentence length, and how much it varies
  - paragraph length and rhythm
  - vocabulary level, and any recurring word choices
  - how arguments are opened and closed
  - use of first and second person
  - what this writer never does

Write that analysis as a style guide of no more than 200 words.
Stop there and show me. I will correct it before you write anything.

SAMPLES:
"""
[sample 1]
"""
"""
[sample 2]
"""
"""
[sample 3]
"""

Once corrected, store the guide in a Project's custom instructions. That guide, not the samples, is what keeps later drafts consistent, and it is portable across chats.

#Controlling register and length

Length instructions work best as budgets applied to parts. “800 words” gets loosely approximated; “four sections of no more than 200 words, one example each” gets followed. Register works the same way: name the reader and their situation instead of asking for a tone.

#Template 4 — summarising with constraints

Summarise the document below for [reader] who needs to [decision].

Constraints:
  - exactly five bullets, maximum 25 words each
  - lead each bullet with the consequence, not the topic
  - include any number that changes the decision; omit all others
  - one final line: what the document does NOT tell us that we would
    need to know

Preserve hedging. If the source says "may" or "preliminary", do not
promote it to "does". If the document contradicts itself, say so
rather than resolving it.

DOCUMENT:
"""
[paste]
"""

“Preserve hedging” is the load-bearing constraint. Qualifiers are the first thing compression drops, and that is how a preliminary finding becomes a fact in a briefing note.

#Template 5 — rewrite for a different audience

Rewrite the text below for a new reader.

ORIGINAL READER: [who it was written for]
NEW READER: [who they are, what they care about, what they distrust,
             how much time they have]
NEW PURPOSE: [what they should do after reading]

Rules:
  - every factual claim survives unchanged
  - define jargon inline on first use, or cut it
  - lead with what this reader has to decide
  - maximum [N] words
  - do not add examples, statistics or claims not in the original

After the rewrite, list separately: anything you cut that the new
reader might still have needed, and anything you could not translate
without losing accuracy.

TEXT:
"""
[paste]
"""

That closing list is what makes this safe on technical material. Rewrites lose things; being told what was lost beats a rewrite that pretends otherwise.

#Why “make it sound human” makes writing worse

It is an instruction with no measurable target, so the model reaches for its idea of human-sounding prose: contractions, rhetorical questions, a chatty aside, an em dash where a full stop belonged. The result reads less like a person than like a machine performing informality. “Make it engaging” and “make it less AI” fail the same way. Replace the vibe with constraints you can check — concrete detail is what makes writing sound like a person, and Claude cannot invent your specifics.

#Which model to write with

Model choice matters far less than the brief. Step up only when the task is hard reasoning dressed as writing — an argument that has to hold together, or a technical explanation that must be exactly right.

Thinking is on by default on Sonnet 5, Opus 5 and Fable 5. It helps with structural decisions, does little for line-level prose, and is billed as output either way. The model comparison covers trade-offs per workload; the API page covers cost control at volume.

#Ownership, disclosure and the limits

You own what you produce. Anthropic's Consumer Terms assign users “all of our right, title, and interest—if any—in Outputs”, and the Commercial Terms say the customer owns its outputs. Commercial use is permitted on every plan including Free; the one universal prohibition is training a competing model on outputs. Clause detail and the retention differences between consumer and commercial plans are on commercial use rights; organisation controls are under Claude for business.

Ownership is not copyright, and it is not disclosure. Whether AI-assisted text is copyrightable depends on jurisdiction and on how much of it is yours. Disclosure is a matter of policy: many publishers, journals, universities and clients now require it, and some prohibit AI assistance outright. Check the rule that applies before you file, not after.

The honest summary: Claude is strong at structure, compression, second drafts and diagnosing what is wrong with a piece. It is mediocre at first drafts from thin briefs, and it cannot supply the specifics that make writing worth reading. For patterns beyond writing, see prompt engineering and Claude examples.

#Frequently asked questions