Claude.ai Projects — long-lived context with files

What a Project is#

A Project in Claude.ai is a workspace that bundles three things:

1. Custom instructions — a system prompt that applies to every conversation in the Project
2. Project knowledge — files you upload once (PDFs, docs, spreadsheets, code) that Claude reads on every conversation
3. Conversations — a history of chats, all sharing the above

You start a new conversation inside a Project; Claude reads the instructions + knowledge fresh each time, then talks with you. Conversations stay listed in the Project sidebar — you can pick one back up days later and Claude still has the same context.

The four real use cases#

1. “My ongoing work on X”#

Best example: a Project called “Writing my novel” with instructions like “You’re my writing partner. Voice is dry · British · understated. Don’t be enthusiastic. Critique honestly. No marketing language.” Plus all the chapter drafts as files. Every chat in the Project inherits the voice + has the manuscript context.

Other versions: “Studying for SC-500,” “Tax planning 2026,” “This codebase refactor” — anything you’ll come back to over weeks.

2. “A grounded reference on a domain you keep coming back to”#

⚠️ Data-classification note (especially for Sush at Microsoft): customer engagement detail, internal-classified material, and anything from WorkIQ / internal Teams / SharePoint does NOT go into Anthropic Projects. Use the public-source slice only — your own notes, public docs, things you’d be comfortable seeing in a public retention window.

Particularly good when there’s a stable, public-source knowledge base. Drop the docs (style guide, public product spec, the textbook you’re working through), set instructions like “This is for studying SC-500. Cite the Microsoft Learn study guide section when relevant. Don’t invent exam questions.” Every chat inherits that grounding.

Avoid this pattern for: anything client-specific, anything internal to your employer, anything you wouldn’t paste into a public document.

3. “My reference shelf”#

A Project loaded with reference material you query often — your company’s style guide, a technical spec you’re learning, a textbook. Treat Claude as a librarian who’s read all of it.

4. “My team’s shared workspace” (Team / Enterprise tier)#

On Team and Enterprise, Projects can be shared across the workspace. Useful for “all of engineering uses this Project for runbook lookups.” Members can read; admins control who can edit instructions / files.

Setting one up well#

The instructions matter more than you think. A vague “You help me with my novel” gets you vague help. A specific “You’re a hard-nosed editor. Critique structure, dialogue rhythm, and pacing. Don’t praise. Suggest cuts.” gets you sharp help.

Worth including in custom instructions:

• Voice / tone — how should Claude speak (concise · honest · British English · technical · etc.)
• What NOT to do — “no marketing language,” “no emoji,” “don’t start replies with ‘Great question’”
• Domain hints — “I work in cloud security · Azure Copilot specifically · NZ context”
• Output preferences — “tables when comparing, lists when listing, prose only when narrative makes sense”
• Tools / commands — if you give Claude specific commands (“/critique,” “/diff,” “/checklist”), define them in instructions

Project knowledge — what to upload:

• Reference docs that don’t change much (style guides, specs, manuals)
• Past work Claude should know about (previous drafts, prior conversations summarised as a doc)
• Datasets Claude will query repeatedly

What NOT to upload:

• Anything sensitive (the threat model is “Anthropic’s server has this file in its retention window”)
• Massive single files (chunk them; a 500-page PDF makes every conversation slow)
• Frequently-changing data (you’ll re-upload constantly and stale versions will leak in)

Limits to watch#

Limit	Where it bites
Project knowledge size	Each tier has its own cap (Pro is generous for individuals; Team/Enterprise raises the ceiling). Anthropic doesn’t publish exact byte/token figures — check the limits in your account settings as they change
Number of conversations per Project	Projects with many conversations get slow to navigate; the UI starts feeling cramped well before any hard limit
File types	Common docs (PDF, txt, md, source files, spreadsheets) work. Verify before relying on it — the supported list evolves
Per-file size	Files larger than the per-file limit need splitting; the limit isn’t published as a single number, check the upload dialogue for the cap

When Claude can’t fit all your Project knowledge into a single response’s context, it selects what looks relevant to the current question. You’ll sometimes notice — Claude misses a detail that’s in a file you uploaded. The fix: either trim less-used files, paste the relevant bit directly into the conversation, or narrow the scope to a single document for that chat.

When to use a Project vs a one-off chat#

One-off chat: the question stands alone, no follow-up expected, no files needed beyond what fits in one prompt.

Project: any of the following are true:

• You’ll come back to this topic over multiple sessions
• There are reference files you’d otherwise paste every time
• The instructions are non-trivial and you don’t want to rewrite them
• You want conversations grouped together for later browsing

The cost of creating a Project is essentially zero — over-create rather than under-create. Worst case, you delete it later.

Common pitfalls#

Symptom	Cause	Fix
Claude ignores instructions partway through a conversation	Conversation got long; instructions fell out of context	Use shorter chats, or use /clear in long Projects, or paste key instructions inline
Newly uploaded file isn’t reflected in answers	File not yet indexed; conversation pre-dates the upload	New conversation, OR ask “what files do you have access to?” to force a re-read
Multiple files repeat the same info; Claude gets confused	Versioning of overlapping docs	Delete old versions; keep one source of truth per topic
Project knowledge “feels stale”	Files were updated locally but not re-uploaded	Re-upload; the file content is snapshotted at upload time
Hit the knowledge limit	Too many big files	Trim or chunk; keep references-of-references rather than full docs

What to do next#

• §WEB.3 Skills feature — custom callable tools across conversations
• §WEB.1 Claude.ai overview — the wider surface
• §API.1 Claude API overview — when you outgrow Projects and want programmatic control