How to build an author-approved voice profile in Claude Projects

A reusable writing voice starts with the author's actual work and decisions. Claude can help organize that evidence, but the author or editor must decide which patterns deserve to become rules. This guide uses a Claude Project to keep an approved profile available to new chats in that project, then checks a fresh draft for factual accuracy and voice. The product steps follow Anthropic's current Projects guide, checked on October 19, 2026; the evidence and approval method is BIG CHANGE's editorial procedure. We did not run an author interview, upload files or test a draft for this article.

The big change

• What changed: Project knowledge and instructions let an author or editor carry an approved writing profile into later chats in the same Claude Project. A previous chat alone is not the documented way to load that profile.
• Why it matters: A small writing team can make its style rules traceable to permitted samples and the author's own answers. That gives the person approving copy something specific to accept, correct or reject, while each draft's facts still need a separate check.
• What to watch: The decisive handoff is the saved profile revision: which evidence supports each rule, who approved it, and whether the next project chat has the right file and instructions. Claude's claim that it used a revision does not establish what was actually saved.

Before you open a project

You need a signed-in Claude account, permission to analyze and upload the writing, and the author's real answers to a short interview. Keep a separate, owned factual assignment with approved source facts for the final check. Choose an intended audience and channel, such as the actual newsletter or product page the author writes for; the rules may differ between channels.

Anthropic says Projects are available on Free accounts, which can create up to five. Plan usage limits still apply. Check your account's current controls and any organization policy before adding material. This guide follows the current Claude chat Projects interface; a separately rolling-out Projects beta is not required.

Review the account's privacy settings and your organization's data policy before uploading writing or interview answers. For consumer accounts, Anthropic documents Settings > Privacy > Help Improve our AI models in its model-improvement settings guide. Choose the setting allowed by the material's owner; turning it off does not undo already-started training and has documented safety-review exceptions. That choice is separate from project sharing and memory. Upload only material the author or rights holder permits.

1. Inventory writing you may use

Choose genuine samples whose author and reuse permission you can establish. For each, record a source ID, author, date, channel, audience and any restriction on uploading or retaining it. Remove confidential passages that the owner has not cleared. Exclude ghostwritten or third-party work unless the relevant rights holder explicitly permits this use. Set aside the separate factual assignment; a later check is more useful when it is not one of the profile-building samples.

Check: Every proposed sample has a known source and permission. If either is uncertain, leave it out. Do not fill a thin evidence set with invented copy.

2. Interview the actual author

Ask whom they address, how direct or formal they want to sound, which words and structures they prefer, what changes by channel, and what they would never say. Record their actual answers with a date and attribution. Keep exact quotations apart from your paraphrases. Mark unanswered or conflicting questions needs owner decision; Claude cannot answer them on the author's behalf.

Expected result: A small evidence set containing permitted writing and attributed answers, with open decisions visible. There is no required number of samples or universal interview script.

3. Create the Claude Project and control its scope

In Claude, open Projects, choose + New Project, and name the project. In its knowledge area, use + to add the permitted text files, retaining their source IDs and rights notes. Anthropic says a project's name and description do not themselves become Claude context. Project knowledge is the documented way to make those files available across chats in that project.

Before uploading to a Team or Enterprise project, check its visibility, members and permissions. If you want to keep this check bounded from account memory, the current Projects guide describes turning off Use account memory under the project's Memory settings. It also describes turning off Memory in each new chat's + menu before the first message. Those controls concern memory; they do not remove knowledge or instructions you explicitly add to the project. If your interface lacks a control, consult the current help page before proceeding. If the project cannot be created or the account has reached its limit, retain the local inventory rather than moving sensitive material to a different account.

4. Turn evidence into a candidate, then audit it

In a chat inside that project, ask Claude for candidate rules in a reviewable table: the exact quoted passage and source ID; the observed pattern; any inference; a counterexample or channel exception; and needs owner decision where evidence is missing or inconsistent. Ask it to organize the proposed profile around audience, tone, language, structure, allowed variation and never-do rules. This is a suggested editorial request, not a tested extraction feature.

Compare every quoted passage with the original file or interview record. Remove invented quotes and claims about the author's beliefs. A repeated pattern can support a proposed rule; it cannot prove the author's intention. Keep exceptions where the author deliberately writes differently for a different audience or channel.

Check: Each rule has a traceable source, a stated inference where necessary, and a real person to settle disputed choices. If evidence conflicts, leave the rule unresolved for the author instead of letting the model choose.

5. Approve and load one exact revision

The author or responsible editor accepts, changes or rejects each candidate rule. Save the approved profile with a revision identifier, date, approver, source IDs and a few permitted examples of preferred and avoided wording. Record unresolved choices separately; do not quietly turn them into instructions.

In the project, open Set project instructions, add the approved reusable rules and their profile revision identifier, then select Save instructions. Add the approved, source-attributed profile file to project knowledge with +. Inspect both saved surfaces to confirm the exact revision identifier is there. Anthropic documents knowledge and instructions as project context for chats in that project. An interview chat by itself does not establish cross-chat reuse, and these instructions do not apply globally to other projects.

6. Check a different factual draft twice

Start a new chat inside the same project. If you are using the bounded memory check from step 3, turn off Memory in that chat's + menu before sending its first message. Inspect the saved profile file and project instructions yourself. Supply the reserved owned assignment, its approved facts and the target channel, then request a draft. You can ask Claude which profile revision it used, but treat its answer as a claim to compare with the saved project content.

Check every material statement against the approved facts and source material: additions, omissions, altered meaning and unsupported claims need correction. Only then check wording against the approved voice rules, examples, audience, channel exceptions and prohibited phrases. Record the two decisions separately and let the author or editor decide whether the draft is usable. A familiar tone cannot make an unsupported fact true.

If the revision appears missing, confirm that the chat is in the intended project and that both the file and instructions were saved there. If a rule misreads a quote or the author rejects it, revise the profile with a new identifier, retain the earlier revision and repeat the affected checks on owned input. Repair factual errors against the approved facts before revisiting style. Recheck human approval when the author, audience, channel or positioning changes.

Limits and reuse

Anthropic's upload guide says each project file can be up to 30 MB and the project's knowledge must fit its context window. Anthropic's Projects overview says paid plans can expand large project knowledge through RAG; that is optional for this small-text procedure. Usage limits and paid-plan costs vary, so check the account before a large upload or repeated drafts. No fixed completion time, fidelity score or productivity gain follows from these documentation-only steps.

If your goal is to run a repeat review of drafts you already have, our Gemini Skill writing-review guide covers that different task. Here the output is an author-approved profile, followed by a checked draft on new factual input.