Official sources reviewed · October 3, 2026
If Claude rejects a file, identify the destination before changing the document: a normal chat, Project knowledge, and a code-execution workflow are not interchangeable. Also separate an attachment that fails to arrive from one that arrives but produces an incomplete answer.
The checks below are a suggested diagnostic workflow, not a claim that HiseHub tested every Claude plan. Save the exact message and keep the original file unchanged. One small, non-sensitive control file can give you more useful evidence than repeatedly submitting the entire batch.
Claude file limits: which official rule applies?
| Surface or content | Published guidance | How to use it |
|---|---|---|
| Ordinary chat attachment | Upload guide: 500 MB per file, up to 20 files per chat | Not a promise for every file-processing tool |
| Project knowledge | 30 MB per file; no fixed file-count cap listed | Capacity and retrieval still matter |
| Images | Up to 8000 × 8000 pixels | Dimensions and file weight are different checks |
| PDFs | At most 1000 pages; visual analysis through 100 pages, text-only for 101–1000 | Count PDF pages, not printed labels |
| Code execution / file creation | Separate guide also lists 30 MB for uploads/downloads | See the source discrepancy below |
Supported examples include PDF, DOCX, TXT, CSV and common images. XLSX requires code execution/file creation. Non-PDF document uploads extract text rather than embedded pictures. Claude upload guide.
The official pages are not fully consistent: the upload guide gives the ordinary-chat figure above; the file-creation guide publishes the smaller figure and describes a computing path for larger PDFs. They do not establish one universally guaranteed limit for every workflow. Follow the error on your actual account and test a small copy; do not assume a subscription upgrade repairs the discrepancy.
Run a controlled test before converting everything
- Prepare a harmless TXT control. Write two short sentences with a unique label and a number you can check. Do not use company or personal information.
- Use a new ordinary chat. Attach only the control and ask Claude to repeat the label and number. Record whether the attachment is rejected, remains pending, or is accepted.
- Try one affected file. Use the same account and destination. Do not change the browser, model, network and file simultaneously.
- Change one thing next. A newly exported copy tests document structure; a new conversation tests conversation state; an approved alternative client tests the client path.
| Observed result | What it suggests | What it does not prove |
|---|---|---|
| The TXT control works; one PDF fails | Inspect that PDF’s structure, encryption, pages and extraction | That all PDFs or the whole account are broken |
| A chat attachment works; Project knowledge fails | Investigate the Project destination and its capacity | A general connection outage |
| All small controls fail across approved clients | Check account access and service status; preserve evidence for support | That reducing a large file will solve it |
| Attachment succeeds; the answer misses a chart | Test content visibility and a narrower question | That the network upload failed |
PDFs: the exact 100-page boundary matters
The PDF rules in the table include the boundary itself: a 100-page document is not excluded merely because it is not “under 100.” A longer accepted document can have different visual processing. If your task depends on a diagram, preserve a focused PDF excerpt or an authorized image of that diagram rather than silently replacing the whole report with plain text.
Use both references when a report has a cover, Roman-numbered front matter or restarted page labels: “PDF viewer page 12, printed page 7.” Check the answer against that page. A plausible summary is not evidence that every row, footnote or qualification was read.
For a text-heavy PDF, try selecting a sentence in your local viewer and copying it into a local text editor. Check reading order, units, minus signs and column boundaries. Empty or scrambled text may call for a fresh export or an authorized OCR workflow. Do not remove another person’s copy protection or passwords.
Inspect the extracted text locally
AI File Preflight checks PDF/TXT/Markdown locally, flags incomplete or low-text pages and creates budget-sized text chunks. It does not OCR, preserve chart layout, use Claude’s tokenizer or test your Claude upload allowance. Its resource caps belong to HiseHub, not Claude.
Use the original PDF alongside any exported chunks. Read the first and last lines of each chunk, keep headings with their exceptions, and carry source-page references into your questions. Splitting can lose cross-section meaning even when each text part fits your chosen budget.
Projects: storage capacity is not “everything in context”
Claude’s Project RAG documentation describes automatic retrieval on paid plans as knowledge approaches context capacity, expanding stored knowledge by up to 10×. Relevant passages are searched instead of all documents being loaded together. The old instruction that every Project must fit wholly in one context window is therefore incomplete.
For an important answer, name the intended source and revision, request a supporting passage, and check it in the original. Uploading two conflicting policy versions without telling Claude which is current can create an evidence problem even when retrieval works. Remove obsolete material only after retaining what you need and checking Project sharing implications.
Separate length limits from a usage-limit warning
Claude distinguishes conversation length and usage allowances. Usage depends on the work performed; a response reserve also reduces usable context. A named local token count is not your Claude usage meter. For a long-chat problem, follow Claude conversation too long; for a reset warning, use the session-limit guide.
Waiting may help a time-based allowance; it does not repair a malformed file. A smaller excerpt may help a length problem; it does not prove that missing subscription access has returned. Keep those diagnoses separate.
When to stop retrying and contact support
Check Claude Status before another upload during a broad failure. Gather the exact error, time and timezone, account/workspace surface, client version, file extension, byte size and PDF page count. Explain whether your harmless control worked and which single change altered the outcome.
Do not send credentials, private documents, unredacted logs or customer information merely to make a support request look complete. Use an authorized support route. Do not disable organizational security controls or clear an entire document library as an experiment.
HiseHub is independent of Anthropic. Sources reviewed October 3, 2026; live account errors and provider changes can alter behavior. The official limit discrepancy is disclosed, not resolved through an invented universal figure.
Working across multiple chats: Learn how to use Claude Projects, project knowledge, files, and instructions.