Use Claude Projects for reusable sources, not as an unverified archive. Put approved reference material in project knowledge, keep recurring behavior rules in project instructions, and use a separate chat for each task. Before relying on an answer, check that it comes from the correct document and version.
This guide covers the existing chat-project workflow at claude.ai/projects. Anthropic is also rolling out a different Projects beta, initially for selected Pro and Max Claude Code users. Its parallel-thread workflow is not the interface described below; existing chat projects continue to work during that rollout. See the official Projects overview for that distinction.
Can free users use Claude Projects?
Yes. Anthropic currently allows free users to create up to five projects. Enhanced project knowledge with retrieval-augmented generation (RAG) is a paid-plan feature, available on Pro, Max, Team and Enterprise. Organization sharing belongs to Team and Enterprise and can be restricted by administrators. A missing RAG or sharing control therefore does not prove that ordinary Projects require payment. Source: Projects availability.
Set up a project without mixing sources and instructions
- Open Projects and choose New Project. Use a name that helps you find the workspace later.
- Add the reusable documents to project knowledge. Put only the behavior rules in Set project instructions.
- In an organization workspace, check visibility before adding material.
- Start a new project chat with one small task and an evidence check.
The project name and description are organizational labels, not task context. Important requirements belong in the instructions or source files. Interface labels can move; follow the corresponding controls in your account. Source: create and manage projects.
Choose the right place for each piece of information
| What you have | Recommended location | How to check it |
|---|---|---|
| An approved reference, requirement or decision | A clearly versioned knowledge file | Ask for the file name, section and supporting evidence |
| A recurring format or behavior rule | Project instructions | Give one small task with an observable pass/fail condition |
| A temporary draft or error log | The current task chat | Include only the part needed for that task |
| An important conclusion from a previous chat | A source-backed decision log | Verify the conclusion before saving it for reuse |
This is HiseHub’s suggested organization method, not an official storage guarantee. The point is to make an answer auditable: “this is the current source” is more useful than “Claude probably remembers it.”
Project knowledge, memory and chat history are not interchangeable
Do not assume that adding a file to one chat creates a shared, version-controlled source for every other chat. Project knowledge is the explicit reusable source layer. Claude’s separate memory feature can also carry context forward, and Anthropic describes a dedicated memory space and summary for each project. Whether memory is used depends on account and chat settings. Source: chat search and memory.
The project-management help page still says cross-chat context must be added to knowledge, while the memory documentation describes an additional continuity mechanism. Our practical reading is to avoid an absolute “Claude can never remember another chat” claim, but still keep important decisions in an explicit source. Memory is not evidence that a particular attachment, exact quotation or latest revision was used.
Project file limits and paid RAG: two separate boundaries
The upload guide lists 30 MB per Project file, no fixed Project file-count limit, and a content/context constraint. Those are not the same rules as an individual chat upload. Non-PDF documents use text extraction; XLSX requires code execution and file creation. Check the official upload guidance and our Claude upload troubleshooting guide for format and PDF boundaries.
For eligible paid projects, RAG activates automatically as knowledge approaches the context boundary. Anthropic describes up to 10 times more knowledge capacity and a visible RAG indicator. Retrieval selects relevant passages instead of putting every file into every request. This does not multiply the model’s response context by ten, promise exhaustive reading, or remove account usage limits. Source: Project RAG.
A five-check method for verifying document retrieval
Use a small, non-sensitive document whose contents you know. The procedure below is an original reader exercise, not a claim that HiseHub tested your account or measured Claude’s accuracy.
- Make the expected answer first. Choose one factual detail near the start, one in the middle and one near the end. Record the correct source and section for each.
- Ask for evidence, not confidence. Name the file and request the answer plus the smallest supporting passage. Compare the returned wording with the actual source.
- Test a missing fact. Ask a question the source cannot answer. A useful response identifies the gap instead of inventing a project fact.
- Check the version. If two approved revisions exist, specify which is authoritative and ask for its revision date. Keep an archive outside the active knowledge set when it is not needed.
- Repeat the important checks in a new project chat. Record whether the same facts and source version are recovered. One correct answer does not establish reliable retrieval across a large collection.
For example, create a tiny fictional product brief with three fields: audience, supported export format and a deliberate “not yet decided” delivery date. You know all three answers before asking Claude. If it invents a delivery date or uses an obsolete export format, narrow the source set and retest before adding real work.
A project-instructions template you can adapt
Keep this short. Replace the bracketed fields; do not paste the entire source archive into the instruction box.
Goal: [the recurring task]
Audience: [who will use the result]
Approved source: [file name and revision]
Evidence: identify the source section for factual claims.
Missing information: say what is missing; do not invent an answer.
Conflicts: show conflicting sources and ask which is authoritative.
Output: [the required format]
Sensitive data: do not request or reproduce credentials or private identifiers.Blank template only. Fill it in privately; remove personal data before sharing.
A good acceptance check is specific: “the answer names the current source and flags the unresolved date.” “Be smart and accurate” gives you little to verify. Save the successful prompt and expected evidence alongside your decision log so the next review is repeatable.
When a file is accepted but the answer is wrong
| Symptom | Next diagnostic step | Avoid |
|---|---|---|
| The wrong revision appears | Name the authoritative file and inspect the supporting passage | Assuming a more expensive plan fixes source conflicts |
| A scanned page or chart is missed | Check whether usable text or the relevant visual is actually available; provide a readable excerpt | Calling a successful upload proof that every page was understood |
| A large knowledge set gives vague answers | Repeat one known-answer question against a narrow source set | Adding more files before isolating the failure |
| Only one old chat fails | Preserve the useful state and try a focused continuation | Deleting the original chat before saving what matters |
| Small new chats also show a usage warning | Read the account’s Usage meter and reset time | Repeatedly opening projects to bypass a usage allowance |
These are diagnostic branches, not claims that we know the cause in your account. Context length and account usage are separate limits. Read Claude conversation too long for a compact continuation brief, or Claude usage-limit resets when the meter is the problem. Source: usage versus length limits.
Check extracted text before sending it
The AI File Preflight Beta can inspect PDF/TXT/Markdown text locally, flag low-text pages and produce smaller TXT chunks. It does not use Claude’s tokenizer, verify a Project allowance, perform OCR, preserve PDF visuals or predict account usage. Its named text encodings are preparation aids, not Claude token counts. Keep the original source and page references when using an extracted chunk.
Before using a project for real work
- Confirm authorization to process the material with the selected account and workspace.
- Keep an independent copy of the approved source and decision log.
- Check the project visibility and memory settings that apply to your work.
- Record important factual checks; a fluent answer is not an acceptance test.
- Repeat the small source checks after replacing files or changing the workflow.
If you are deciding whether paid RAG is useful, first try the known-answer exercise with your normal source set. Compare Claude plan differences and regional App Store prices only after identifying the feature or allowance you actually need. This guide does not claim that upgrading fixes inaccurate retrieval.
Sources and review scope
Reviewed October 3, 2026 against the official sources linked beside the relevant claims. The existing chat-project workflow, free-plan access, paid RAG boundary, memory distinction and Project file rules were checked. This is source-based guidance with an original verification exercise, not an account-specific product test or a retrieval benchmark. Beta rollout, workspace controls and interface labels can differ.
More practical guides: Claude Help · How-To library · HiseHub editorial policy.