Claude Projects Prompts

A Claude Project is a persistent workspace inside Claude.ai. It bundles custom instructions (a system prompt that runs on every chat in the Project), Project knowledge (files Claude can cite), and a chats list scoped to the Project. A regular chat is one-off. A Project keeps the instructions and knowledge wired into each fresh conversation you open under it, which removes the need to keep re-pasting the role. Projects launched on Claude Team on June 25, 2024 and expanded to Pro on September 30, 2024.

Yes. Projects requires a paid plan. Pro is $20 per month for solo users (verified on anthropic.com/pricing on May 26, 2026) and gives you Projects with custom instructions, file uploads, and the Sonnet 4.5 daily quota. Max at $100 or $200 per month raises the quota and unlocks heavier Opus 4.5 use. Team at $25 per seat per month adds shared Projects across a workspace. Enterprise adds SSO, audit logs, and longer retention windows. Free users do not have Projects.

Each Project has a context budget of 200K tokens of conversation plus the file storage that holds your Project knowledge. The practical ceiling is the moment Claude can no longer fit all of the chat history plus the knowledge it needs to reason about into the 200K window. For most Projects, that means 30 to 60 documents of typical length is fine. Past that, Claude starts paraphrasing instead of quoting and you should split into multiple smaller Projects.

The custom instructions hold the persistent context that does not change between chats: role, audience, voice, output format, citation rule. The prompt holds the specific task of the moment. A good test is whether you would copy and paste the same sentence into every new chat. If yes, it belongs in the custom instructions. If it changes per chat, it belongs in the prompt. The single most common Projects mistake is re-explaining the role in every prompt.

Under 300 words is the practical sweet spot. Past that, the instructions start contradicting themselves and Claude adheres to whichever rule it can satisfy. The pattern that worked in our testing was four sentences: one role sentence, one audience sentence, one format rule, one citation rule. Add specifics in the Project knowledge instead of stuffing them into custom instructions. We tested 14 custom-instructions sets of varying length and the short ones outperformed the long ones on 11 of 14.

Mostly yes, with caveats. Files are retrieved by relevance to the current prompt and surfaced to Claude in the 200K-token context window. For small Projects (a handful of files), Claude reads everything. For large Projects (40-plus files), Claude retrieves the most relevant chunks. The implication is that you should write prompts that name the document or topic so Claude pulls the right material. A vague prompt against a large knowledge set produces a generic answer.

Artifacts (launched June 20, 2024) work the same inside Projects but they persist across chats. An Artifact you created in chat 1 stays available for chat 2 to reference and update. This is the killer combination for iterative work: a long-running React component or a Markdown doc that lives across many chats. Use Projects plus Artifacts together when the deliverable grows over time. Use a regular chat for one-off Artifacts you do not need to revisit.

Yes, on the Claude Team and Enterprise plans. A shared Project lets every member add to the knowledge and start new chats under the same custom instructions. Solo Pro Projects stay solo. Shared Projects are the closest Claude.ai gets to a collaborative workspace and the right place for cross-team workflows like brand voice, runbooks, or vendor reviews. Personal Projects can be exported as Markdown if you need to hand a workstream off without granting access.

Claude Sonnet 4.5 is the daily default. It hits the best mix of quality, speed, and quota for almost every Project chat. Reach for Claude Opus 4.5 when a chat needs to reason carefully across a very long uploaded document or to tackle the hardest writing tasks (expect tighter quota). Reach for Haiku 4.5 on cheap fast work where Sonnet would be overkill. The model picker is per chat, so a single Project can route different chats to different models as the work demands.

Three common reasons. First, the user message in the chat overrides the instructions (a request to answer briefly wins over an instructions line that asks for length). Second, the instructions are too long and self-contradictory (cut past 300 words). Third, a file in Project knowledge that Claude is citing argues against the instructions and the model defers to the source over the format rule. The fix is shortening the instructions and labeling the knowledge files by role so Claude knows which authority to pick.

Not directly inside Claude.ai. There is no built-in versioning on custom instructions or knowledge files. The workaround we use is to keep the custom instructions as a Markdown file in our own version control and update the Project setting by pasting from the source of truth. The same applies to knowledge files: store the canonical version in your own system, upload the latest into the Project. Treat the Project as the runtime, not the repo.

Treating a Project chat as if it were a stand-alone conversation and stuffing the entire brief into the opening message. Custom instructions already carry persistent context. Project knowledge already carries the documents. Each chat is best treated as one tight ask. Newcomers tend to write a 400-word opener that restates role, audience, and format rules already encoded in the settings. Trimming the chat opener down to the real question almost always raises output quality.