Claude Projects: How to Use Persistent Workspaces for Long-Term Work
Claude Projects are persistent workspaces in claude.ai that keep your custom instructions and uploaded files available across every conversation within that project. Unlike a standard chat where context resets each session, a Project holds your role definition, constraints, reference documents, and communication preferences — so Claude starts every new conversation already knowing your stack, your style, and your goals. Projects are included in Claude Max subscriptions and are the recommended way to work with Claude on anything that spans more than one session.
What Claude Projects actually are
Every standard Claude conversation is isolated. When you close the tab and start a new one, Claude has no memory of what you discussed before. Projects solve this for ongoing work.
A Project is a named container that holds two things permanently:
- Custom instructions — a text block Claude reads at the start of every conversation in the project. Think of it as a persistent system prompt you write once.
- Uploaded files — up to 30 files (PDFs, code files, CSVs, text files, Markdown) that Claude can reference throughout every conversation in the project.
Start a new conversation inside a Project and Claude already has your instructions and files loaded. You skip the re-introduction every time.
How to create a Project
- In the claude.ai left sidebar, click New Project.
- Give it a name that reflects the domain (e.g., "SaaS Backend", "Book Draft — Chapter 1-10", "Client: Acme Corp").
- Click Edit instructions to add your custom instructions.
- Upload any reference files using the Add content button.
- Start a new conversation — it will inherit everything you just configured.
You can edit instructions and add or remove files at any time. Changes take effect in the next conversation you open in that project.
What to put in custom instructions
Custom instructions are a short text block — aim for under 2,000 words, ideally under 800. They should cover:
- Role: what Claude is helping you with
- Technical context: the stack, tools, frameworks, or domain-specific vocabulary
- Communication preferences: response length, code style, tone, language (English, Korean, etc.)
- Hard constraints: what Claude should never do (e.g., never use
anyas a TypeScript type, never suggest a rewrite when a small edit will do)
Example: software developer project
You are helping me build a SaaS product called [product name] using Next.js 15, TypeScript, and Postgres.
Context:
- I'm the solo developer
- We use Tailwind CSS for styling
- Our database schema is in /schema.sql (uploaded)
- Authentication is via Clerk
Preferences:
- Write TypeScript, never plain JS
- Prefer async/await over .then() chains
- Always handle errors explicitly
- Don't use any as a type
Example: content writing project
You are my writing collaborator for a blog targeting senior software engineers.
Context:
- The blog is claudeguide.io
- Audience: developers who know Python/TypeScript and use AI tools daily
- Style guide is uploaded as style-guide.md
Preferences:
- Active voice, short paragraphs
- No motivational filler ("In today's fast-paced world...")
- Code examples should be complete and runnable, not pseudocode
- Flag when I'm writing something the audience likely already knows
Example: research project
You are helping me synthesize research on AI adoption in enterprise software teams.
Context:
- I'm writing a 6,000-word report for a non-technical executive audience
- Source documents are uploaded as PDFs
- Citations should use author-year format
Preferences:
- Summarize claims accurately — never overstate what a source says
- When multiple sources conflict, flag the disagreement rather than picking one
- Keep jargon minimal; define any technical term on first use
File uploads: what works and how Claude uses them
Projects support up to 30 files per project. Supported types include PDFs, plain text, Markdown, CSV, most code file formats, and Word documents.
Claude treats uploaded files as reference material it can read on demand. Practical uses:
- Database schemas — upload your
schema.sqlso Claude knows your table structure without you pasting it each time - Style guides — tone and voice documents, brand guidelines, or editorial rules
- API documentation — upload a reference doc so Claude can write correct API calls without hallucinating endpoints
- Existing codebase excerpts — paste key modules as
.txtfiles so Claude understands your architecture - Research papers or reports — upload PDFs to ask questions, extract data, or request summaries
Upload reference documents, not conversation logs. Conversation logs become stale quickly and bloat the file count. If a document changes, replace the file — there is no versioning, so the latest upload is what Claude reads.
Project memory vs conversation memory: what is different
Claude.ai has two distinct persistence mechanisms and they are easy to confuse.
Project context (this article) stores custom instructions and files you explicitly upload. It is scoped to the project. Every conversation in the project sees the same instructions and files. You control exactly what is there.
Claude.ai Memory is a separate feature that extracts facts from conversations and stores them globally — across all your conversations, not just one project. Things like your name, timezone, preferred coding language, or recurring preferences. Memory is automatic and cross-project; Project context is manual and project-specific.
The practical difference: use Project custom instructions for task-specific, technical, or confidential context you want under your explicit control. Let Memory handle lightweight personal preferences that apply everywhere.
Which plan includes Projects
Projects are available on Claude Max (the paid claude.ai subscription) and on Claude for Work (the team and enterprise tiers). They are not available on the free tier.
Claude Max starts at $20/month (5x message limit) and $100/month (20x message limit). If you are doing ongoing work with Claude — software projects, writing series, client work — the Projects feature alone typically justifies the cost of the subscription within the first week of use.
Best use cases for Projects
Long-running software projects — Solo developers and small teams use Projects to keep their stack, schema, and preferences loaded at all times. No more pasting your database schema into every session.
Writing books or article series — Upload your outline, style guide, and completed chapters. Claude can maintain consistency across drafts without you re-establishing context.
Ongoing research — Upload source documents, define citation format, specify your audience, and interrogate your sources across multiple sessions without losing thread.
Client work — Create one project per client. Upload their brand guidelines, past deliverables, and tone-of-voice documents. Switch between clients without context bleed.
Code review and refactoring — Upload your codebase conventions document and linting rules. Every review conversation starts with the same standards applied.
Tips for effective Projects
One project per context domain. Resist the temptation to put everything in one "work" project. A dev project and a writing project should be separate — the instructions are different and mixing them dilutes both.
Keep instructions under 2,000 words. Long instructions are not better instructions. Claude reads the full block every session; vague length adds noise, not clarity. Write tight, specific instructions and cut anything that does not change Claude's behavior.
Upload reference docs, not conversations. Conversation exports go stale the moment you have another session. Upload durable artifacts: schemas, specs, style guides, reference documentation.
Revisit instructions monthly. As a project evolves, your instructions should evolve too. Outdated constraints ("never use Postgres RLS" from before you enabled it) actively mislead Claude.
Test new instructions with a throwaway conversation. Open a new conversation in the project and ask Claude to describe what it understands about the project. This surfaces gaps or contradictions in your instructions before they cost you in a real work session.
Frequently asked questions
Do Projects work with all Claude models? Yes. Inside a Project, you can select whichever model is available on your plan (Haiku, Sonnet, Opus). The custom instructions and files apply regardless of which model you pick for a given conversation.
Can I share a Project with a teammate? Sharing is available on Claude for Work (team and enterprise plans). On a personal Claude Max subscription, Projects are private to your account. If collaboration is important, look at the team plan or build a shared system prompt at the API layer instead.
How many Projects can I create? There is no published hard limit on the number of Projects. The practical limit is that each Project can hold up to 30 files.
What happens to files I upload — can Anthropic see them? Anthropic's privacy policy applies. For sensitive or confidential documents, review the current policy at anthropic.com before uploading. For highly sensitive material, consider using the API with your own infrastructure rather than claude.ai.
Is a Project the same as a system prompt in the API? Functionally similar but not identical. Custom instructions in a Project behave like a system prompt — Claude reads them before every conversation. However, the API gives you more control: you set a system prompt per API call, while Project instructions are fixed until you manually edit them. The API is also more appropriate when you need programmatic control, multi-user access, or custom memory logic.
Take It Further
Power Prompts 300: Claude Code Productivity Patterns — Section 2 covers Project Setup Patterns: the custom instruction templates for 12 common project types (dev projects, writing, research, client work), file organization strategies, and the system prompt patterns that make Claude immediately useful in any new project.
30-day money-back guarantee. Instant download.