What this covers
Files attached to Claude behave differently from text pasted into a chat — they live in a separate retrieval layer, can be referenced by name across messages, and persist inside Projects. The pain this solves: people who paste the same 30-page PDF into ten chats because they didn’t realize Files exist, or who upload one giant zip and then can’t get a useful answer out of it. Audience: anyone with PDFs, CSVs, contracts, transcripts, or codebases they reference more than once.
Key tools and concepts:
- Claude: Anthropic’s conversational AI with file, long-context, and tool support.
- File attachment: A document uploaded into a chat or Project; Claude indexes it and references it on demand.
- Project files: Files attached to a Project, available across every chat in that Project without re-uploading.
Who this is for
People with PDFs, CSVs, or documents they reuse — research papers you cite repeatedly, contracts you compare against, transcripts you mine, style guides you want every output to follow.
When to reach for it
Whenever you would otherwise paste the same document into many chats — that is the signal to move it into a Project. Also when a single document is too large to paste comfortably (3-4 pages or more), or when you want to compare multiple files side by side.
Before you start
- Rename files to something Claude can reference by name. “contract-vendor-a-2026.pdf” beats “Final_v3 (1).pdf”. The agent will quote the filename in answers; readable beats cryptic.
- For multi-file work, decide if they belong in one Project (recurring use) or one chat (one-off comparison).
- Estimate the file’s effective length. PDFs with images or scanned text often have more visible content than text content; expect OCR noise.
- Check your plan’s file size limits and supported types before you spend time uploading.
Step by step
- Upload files to a Project for reuse, not just one chat. Click “Add content” in the Project sidebar, choose Files, drop them in. They stay across every conversation in that Project.
- Ask Claude to read in chunks for very long docs. “Read pages 1-20 first and give me the table of contents” beats “summarize this 200-page PDF” — long-context recall in the middle degrades silently.
- Reference specific filenames in your prompts. “In contract-vendor-a-2026.pdf, find the termination clause and quote it” is much more reliable than “find the termination clause”.
- Ask for direct quotes with page numbers for any claim that matters. “Quote the relevant passage and the page or section header” turns Claude from confident paraphraser to citation-aware reader.
- Delete files when done — they count toward Project storage limits. Keep current contracts, drop obsolete drafts.
- For files you will edit, keep the canonical version in Drive or Git. The Project copy is a working snapshot, not source of truth.
First-run exercise
- Create a Project called “PDF practice”. Upload 1 paper or contract you have read before so you can spot hallucinations.
- Ask: “Describe this file’s structure — sections, length, and any tables or figures. Do not summarize content yet.”
- Drill into one section: “Quote the key claim in section 3 and the supporting evidence.” Verify against your prior reading.
- Start a second chat in the same Project. Confirm the file is still available without re-uploading.
Quality check
- Did Claude reference the actual filename / section / page, or paraphrase generically? Generic answers mean it did not really retrieve.
- Spot-check 3 quotes from the answer against the source file. OCR errors and hallucinations both surface here.
- Watch the token-window indicator on huge uploads — when it is full, recall drift increases (Claude long-context unstable).
- Confirm scope: when asked about file A, did Claude blend in file B? If yes, ask follow-ups with explicit filename anchors.
How to reuse this workflow
- Build one Project per long-running need: “Vendor contracts 2026”, “Research papers Q2”, “Internal style guides”.
- Save your standard prompts as a doc in the Project (“Standard questions for any contract: termination, liability, SLA, payment terms”). New chats pick this up automatically.
- Rotate files quarterly — remove outdated docs so Claude does not blend stale and current content.
- Re-evaluate file size limits and supported types each quarter; both shift.
Recommended workflow
Upload to a named Project -> reference per file -> ask chunked questions with page-cite requirements -> verify 2-3 quotes by hand -> clean up files you no longer need. Repeat the workflow weekly for active research.
FAQ
- Can Claude read scanned PDFs?: Yes, via OCR, but quality drops on dense or low-resolution scans. Spot-check quoted passages.
- What about Excel / CSV?: Claude can answer questions over them and output structured Markdown / CSV back; verify totals by hand on important numbers.
- Does the file count against context every message?: No, the retrieval layer fetches relevant chunks per query. Context fills based on how much it pulls.
- Are files private to my Project?: Yes, but anyone with Project access can see them. Don’t share Projects holding sensitive files.
- What is the size limit?: Varies by plan and file type; check the help center as it changes quarterly.
Common mistakes
- Re-uploading the same file every chat instead of using a Project — wastes time and breaks the cross-chat reference habit.
- Asking Claude to summarize without specifying which file — when multiple are attached, it blends sources.
- Loading a giant file once and asking a single broad question — long-context recall in the middle of the document degrades silently (Claude long-context unstable).
- Trusting unsourced claims. Always re-ask for the exact passage and page.
- Leaving stale files in a Project for months — old contracts confuse questions about current ones.
- Treating Claude’s OCR as perfect. Heavily scanned PDFs may have garbled passages — sanity-check.