Claude Project Setup for Writers

Put stable context in the project

A Claude project is most useful when it contains context that should apply across many chats. For a writer, that means the voice guide, house rules, source hierarchy, recurring audience notes, glossary, and current outline.

Do not upload every old draft, transcript, and note unless the project genuinely needs it. Too much context can bury the signal. A curated two-page guide is often more useful than a folder of stale files.

The project instruction should describe behavior, not just taste. Tell Claude how to handle uncertainty, when to ask questions, when to refuse to draft, and how to label unsupported claims.

• Voice guide: examples, patterns, and forbidden cliches.
• Source hierarchy: primary sources first, secondary sources labeled.
• Editorial workflow: diagnose, outline, draft, critique, verify.
• House rules: disclosure, privacy, banned claims, review owners.

Write instructions as operating rules

Account instructions apply broadly. Project instructions apply to the specific project. Use that distinction. Put your general communication preferences at the account level and project-specific writing rules inside the project.

A good project instruction says what Claude should do by default: ask clarifying questions when the brief is incomplete, separate facts from interpretation, preserve intentional voice, and provide a change table for edits.

If a rule is not always true, do not put it in global instructions. Put it in the relevant project, skill, or prompt.

Use skills and artifacts for repeatable work

Anthropic describes skills as reusable instruction bundles. For writers, that means a repeated editorial checklist, brand voice, source-review procedure, or manuscript-continuity routine can become a skill instead of a prompt you paste every time.

Artifacts are useful for substantial standalone content: outlines, tables, drafts, checklists, and reusable templates that need a dedicated surface separate from the chat flow.

The rule is practical: if you use it once, prompt it. If you use it weekly, make it a project instruction or skill.

Know what memory carries between manuscripts

Claude memory changed shape in September 2026, and the change matters for anyone working on something long. The memory used in chat is now the same memory used in Claude Cowork, so context built up in conversation is present when Cowork picks up a task, and what happens in Cowork carries back to chat. Claude also adds topics to memory as you chat rather than summarizing a conversation after it ends.

Memory is not your story bible, and treating it as one is the mistake to avoid. What Claude stores is a set of short topic files about you and your working context, visible under Topics in Memory settings, where each one can be read, edited, or deleted individually. Manuscript canon — character sheets, timeline, continuity ledger, style sheet — still belongs in project knowledge, where you control the exact wording and can see the whole document at once. Memory is the ambient layer; the project is the binder.

The practical risk for writers is bleed. If memory spans every surface and every project, a preference absorbed while drafting one book or one client’s copy can surface while you are writing another. The fix is cheap because the topic files are short and editable: when a voice note, a stale deadline, or an old company name starts showing up in the wrong manuscript, open Memory settings and correct or delete that topic. One correction propagates to every later conversation. Memory can also be paused or reset outright.

Defaults differ by plan, which is worth checking before you rely on any of this. Memory is on by default on Free, Pro, and Max across web, desktop, and mobile. On Team and Enterprise, admins control availability and memory stays off for individual users until they turn it on. Topics that Anthropic treats as sensitive — health, race, ethnicity, religious beliefs, politics, gender identity, and similar areas — are not stored by default; you can opt in under Memory settings, in which case they are saved going forward only and never retroactively. Some categories are never stored even with that setting on, including government identification numbers, criminal history, and immigration status. For a deeper treatment of memory systems, see claudememories.com.

• Memory in chat and Claude Cowork are the same memory as of September 2026.
• Topic files are individually readable, editable, and deletable in Memory settings.
• Keep manuscript canon in project knowledge, not in memory.
• Audit topics when you switch books, beats, or clients to prevent voice and fact bleed.
• Sensitive topics are off by default and opt-in applies only going forward.

Maintain the project like an editorial room

Every project needs cleanup. Remove obsolete drafts. Update the source packet. Move completed work out of Claude. Refresh the style guide when the publication or manuscript evolves.

For teams, nominate an owner for the source hierarchy and house style. Otherwise, Claude may faithfully apply stale rules no one remembers approving.

Maintenance is part of writing quality. The assistant can only follow the context you give it.

1. Review project instructions monthly or after major scope changes.
2. Update source packets when product, legal, or factual claims change.
3. Archive stale drafts outside the project.
4. Keep a changelog of editorial rule changes.