Claude Projects: a persistent workspace for recurring work
If you keep pasting the same background into new chats — the same style guide, the same spreadsheet, the same "here's how we do things" — a Project fixes…
Claude guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on claude.ai
If you keep pasting the same background into new chats — the same style guide, the same spreadsheet, the same "here's how we do things" — a Project fixes that. It's the feature for work you do more than once.
What a Project actually is
A Project is a container around a set of chats that all share the same background knowledge and standing instructions. You add reference files and a few rules once; every conversation you start inside that Project can see them without you re-explaining. A plain chat, by contrast, starts blank every time and forgets everything when you close it.
Projects live on claude.ai (web) and in the Claude desktop and mobile apps. Availability depends on your plan and can change — check the current tiers at claude.ai rather than trusting a number here.
When a Project beats a fresh chat
Reach for a Project when the context is stable even though the questions change.
| Situation | Fresh chat | Project |
|---|---|---|
| One-off question | Best — no setup | Overkill |
| Same reference material reused | Re-paste every time | Load it once |
| Consistent voice or format | Re-explain each chat | Set it in instructions |
| A body of related documents | Doesn't persist | Lives in project knowledge |
| Many threads, one topic | Scattered | Grouped in one place |
A good test: if you keep copy-pasting the same preamble into chat after chat, that preamble belongs in a Project.
How to set one up
- In claude.ai or the app, create a new Project and name it for the job ("Q3 board deck," "support replies," "novel draft").
- Add project knowledge — the files and notes every chat should see.
- Write custom instructions — the standing rules for how Claude should behave here.
- Start a chat inside the Project and work normally.
What to put in project knowledge
Stable reference material: a style guide, product docs, a glossary, past examples of the output you want, a data file you'll ask about repeatedly. Keep it curated — a tight, relevant set beats a dumping ground.
What to put in custom instructions
The behavior you'd otherwise repeat: tone ("plain, no jargon"), format ("always end with a bullet summary"), role ("you're editing, not rewriting"), and hard rules ("never invent statistics"). Short and specific works better than a long essay.
Where it goes wrong
Projects are genuinely useful, but they fail in predictable ways:
- Too much knowledge. Stuffing in dozens of loosely related files can dilute focus and make answers vaguer, not sharper. Prune to what's actually used.
- Stale files. Project knowledge is a snapshot you uploaded — it does not auto-sync with the original. Edit the source spreadsheet and the Project keeps the old copy until you re-add it.
- Still confidently wrong. A Project narrows the context; it doesn't make Claude infallible. It can misread a file or state something false with confidence. Verify anything that matters.
- Instruction creep. Contradictory or overly long instructions can quietly degrade results. Revisit them when output drifts.
When this is the wrong tool
Skip the Project when:
- It's a one-off. For a single question, a plain chat is faster and cleaner.
- The data changes constantly. Live prices, today's calendar, a database that updates hourly — a static snapshot goes stale. A live connection (see MCP connectors) or a fresh paste fits better.
- The document set is huge. Very large or fast-growing archives may be better served by a dedicated search/retrieval setup or a tool connection than by loading everything as knowledge.
- It must be truly private or offline. A Project is still a cloud service. Turning off training does not make a request local — it's still sent to Anthropic's servers. For data that can't leave your machine, a local model on your own hardware is the honest answer, not a Project.
- You need the same setup everywhere. Instructions and knowledge live inside one Project; they don't automatically follow you into other Projects or a plain chat.
One more honest note: a Project is not long-term memory. It's a shared context for the chats inside it, not a running record of everything you've ever told Claude. If you want Claude to carry facts across unrelated conversations, that's a different feature — see the memory guide.
Related
- Files and data analysis — how Claude reads the documents you add.
- Memory — persistence across chats, and how it differs from a Project.
- MCP and connectors — connect live tools and data instead of static uploads.
