Long docs and multi-file work

It is 4:40 p.m. on a Tuesday. Someone drops a 47-page PDF, three CSV exports, and a half-written brief into your chat with the message: “Can ChatGPT summarize this before the 5:00 standup?” You paste the PDF. The reply sounds confident. It invents a metric that is not on page 12. It “quotes” a sentence that does not exist. You almost put the fake number in Slack.

This is Part 3 of the ChatGPT everyday tutorial. Parts 1 and 2 covered a first useful week (write, plan, learn) and better answers with structure and examples. Here we stay in ordinary Chat for a harder everyday job: long documents and multi-file work. You will learn a practical loop (chunk, index, verify), when Projects beat a messy thread, how to refuse secrets, and a multi-file recipe you can reuse without turning every task into Work mode or Codex. If you still mix product names (Chat vs Work vs Codex), skim the ChatGPT product map first. If ChatGPT itself still feels new, start with Learn ChatGPT from scratch.

What you’ll learn

Why dumping a whole book into one prompt fails in predictable ways
A three-step loop: chunk by section, index what exists, verify quotes and numbers
When to use a Project (files + scoped chats) instead of one endless thread
A multi-file recipe: goal, file list, extract list, output format, done check
What never to upload (secrets, credentials, regulated dumps) and how to use redacted samples
Common mistakes that look like productivity and still ship wrong claims

Product limits, file counts, and storage move by plan. Ground truth lives in OpenAI’s Help Center (Projects, file uploads, Library). Treat numbers in this post as orientation. Re-check help pages before you write team policy. Series home: ChatGPT everyday tutorial. Paths also sit on Learn.

Why long docs break naive prompts

ChatGPT can read files you attach or paste. That does not mean “throw the entire annual report in and trust the first paragraph.” Models work with limited attention in a given turn. Long inputs get compressed in ways you do not control. The product may retrieve pieces of a file rather than re-reading every line like a human with a highlighter. When people say “it hallucinated from my PDF,” they often mean one of these failure modes:

Blended pages. A claim from page 4 fuses with a number from page 19.
Plausible filler. The model fills a gap with industry-sounding prose because your prompt asked for “a complete summary.”
False quotes. Quotation marks appear around a sentence that is only close to the real text.
Wrong grain. You wanted “what changed vs last quarter” and got a generic brochure of the whole PDF.
File soup. Three CSVs and a deck get treated as one story without saying which file each claim came from.

None of that means “never use files.” It means long-document work is a process, not a one-shot magic button. The same discipline you use for short structured prompts in Part 2 (role, goal, constraints, example, format) still applies. You just add file hygiene on top.

Rule of thumb: If you would not trust a junior analyst who skimmed for eight minutes, do not trust a one-shot “summarize this PDF” reply either. Give structure, demand citations, and check the quotes.

The three-step loop: chunk, index, verify

Use this loop for any doc longer than a few pages, or for any job that mixes more than one file. It is boring on purpose. Boring is how you avoid the Slack incident.

Three steps for long docs: chunk by section or chapter, index by asking what exists first, verify quotes and numbers,…

1. Chunk by section or chapter

Chunking means you do not ask for the whole world in one breath. You split the work the way a careful human would: by table of contents, by heading, by appendix, by file. For a 40-page policy PDF, work chapter by chapter. For a deck, slide ranges. For logs or exports, one logical table at a time.

Chunking helps in two ways. First, each answer stays closer to the text you care about. Second, when something is wrong, you know which chunk to re-open. “Wrong somewhere in the 47 pages” is not a fixable bug report. “Wrong in Section 4, refunds” is.

Practical chunk sizes for everyday Chat (not a hard product limit, a working habit):

One chapter or major H2 at a time for prose PDFs
One sheet or one clear question per CSV (filters, grain, key columns first)
One decision question per turn (“What does Section 3 say we must do before launch?”) instead of “Tell me everything”

2. Index: ask what exists first

Before you ask for analysis, ask for a map. Call it an index pass. You are not requesting a narrative yet. You want inventory:

What sections or sheets exist?
Which pages mention refunds, SLAs, pricing, or the metric you care about?
What columns are in each CSV, and what does one row mean?
Where are tables vs pure prose?
What is missing that you expected to find?

Here is a prompt pattern you can paste and adapt. Swap the file names and the topic.

You have the attached files only. Do not invent sections or columns.

Task: INDEX PASS only (no recommendations yet).

1) List each file by name and type.
2) For each PDF or doc: list headings / sections you can see, with page or section labels if available.
3) For each spreadsheet: list sheet names, column headers, and a one-line guess of the grain (what one row means). Mark guesses as guesses.
4) Flag anything unreadable (scans, image-only pages, truncated tables).
5) End with: "I have not analyzed content yet."

Refuse any request to invent missing pages.

If the index pass is wrong, stop. Fix the attachment, re-export the PDF, or work from a cleaner file. Analysis on a broken map is theater.

3. Verify quotes and numbers

When you finally ask for summaries or comparisons, demand checkable claims. Prefer short excerpts with page or section pointers over long paraphrases. Then you do the human job: open the source and confirm.

Verification prompt pattern:

Using only the attached files, answer: [specific question].

Rules:
- Every factual claim must include a pointer: file name + section/heading or page if available.
- For any number, quote the surrounding phrase or table row in quotation marks.
- If the files do not contain the answer, say "Not in the provided files" and stop.
- Do not invent metrics, dates, or quotes.
- Separate: (A) what the files say (B) my interpretation (label both).

Your personal verify checklist (two minutes, not optional when the output goes to other people):

Check	What you do	Pass means
Quote hunt	Search the PDF for a distinctive phrase from the reply	Exact or clearly matching text exists
Number hunt	Find the same figure in the source table or sentence	Same units, same period, same filter
File label	Confirm the claim points to the right file	No CSV claim labeled as “from the PDF”
Scope	Confirm the answer matches your question grain	No “whole company” claim from one region sheet
Missing	Look for “Not in the provided files”	Gaps are honest, not filled with inventiveness

If a quote fails the hunt, treat the whole answer as suspect until re-run with tighter constraints. Do not “edit around” one bad number and keep the rest. Bad quotes often travel in packs.

Projects: keep files and chats in one workspace

OpenAI’s Help Center describes Projects as workspaces that group chats, reference files, and instructions for an ongoing effort. That is the product answer to “I keep re-uploading the same five PDFs into random threads and losing the thread of what we decided.”

Use a Project when:

The same file set will matter for days or weeks (policy pack, vendor RFP, quarterly review binder)
You want standing instructions for that effort only (“Cite section headers. Never invent KPIs. Audience is non-finance managers.”)
You will run many small chats (index, extract, rewrite, quiz yourself on the material) against the same materials
You want separation from personal life and other work streams

Stay in a plain chat when the job is a one-off attachment and you will never open those files again. Projects are organization, not magic accuracy. A Project full of secrets is still a secrets problem. A Project with a vague goal still produces vague work.

Project setup checklist (ten minutes)

Name the Project after the outcome, not the vibe (“Q3 vendor comparison,” not “AI stuff”).
Write 5 to 10 lines of project instructions: audience, refuse list, citation rules, output defaults.
Upload the current file set only. Delete superseded drafts so the model does not blend v3 and v7.
Run an index pass as the first chat inside the Project. Save or pin that map if the UI lets you find it later.
Do analysis chats as separate turns with one question each when stakes are high.

File counts and storage caps differ by plan (Help Center articles on Projects, uploads, and Library). If uploads fail, check plan limits and whether the file is an unreadable scan. Re-export text-based PDFs when you can. Image-only scans are harder for any tool that needs clean text.

No secrets: redaction before upload

Part 8 of Learn ChatGPT and your workplace policy still apply. Multi-file work is where people get sloppy because the job feels urgent. Urgency is not a policy exception.

Do not upload or paste:

API keys, passwords, private keys, session tokens
Full customer PII dumps, health data, or regulated datasets your policy forbids
Unreleased financials if policy says consumer ChatGPT is off-limits for that class
Source code with embedded secrets, connection strings, or production credentials
Anything your company labeled “no AI tools” or “approved tools only” if this product is not on the list

Prefer instead:

Redacted samples (fake names, shuffled IDs, truncated tables)
Public docs and already-approved internal templates
Structure-only exports (column headers + three fake rows) when you only need help writing a formula or query shape
Enterprise or workspace products your IT actually approved, when the data class requires it

A useful mental model: ChatGPT is a skilled intern with imperfect memory and no clearance badge of its own. You decide what enters the room. If you would not paste it into a random vendor form, do not paste it here either unless policy explicitly allows that class of data on this plan.

Multi-file recipe (copy and reuse)

When you have two or more files, use a fixed recipe so you do not improvise under time pressure. The diagram below is the spine. The table after it is the filled example.

Multi-file recipe in five steps: goal, files list, what to extract, output format, done check

Step	What you write	Toy example
1. Goal	One decision or deliverable	“Two-page brief: does Vendor A meet our SLA and price caps?”
2. Files list	Exact names + what each is for	`rfp.pdf` (our requirements), `vendor_a.pdf` (response), `pricing.csv` (their quote)
3. What to extract	Bullets only, no essay yet	SLA hours, uptime %, price per seat, exclusions, open questions
4. Output format	Shape for humans	Table: requirement \| vendor claim \| file pointer \| pass/fail/unclear
5. Done check	How you will verify	Every row has a pointer; no invented numbers; list gaps

Full multi-file prompt skeleton:

CONTEXT
Goal: [one decision or deliverable in one sentence]
Audience: [who reads this]
Refuse list: inventing numbers, inventing quotes, blending files without labels

FILES (use only these)
1) [filename]: [why it is here]
2) [filename]: [why it is here]
3) [filename]: [why it is here]

PROCESS
Step A: Index pass (sections/columns only).
Step B: Extract the fields listed below. One bullet per field with file pointer.
Step C: Build the output format. Mark unclear items as UNCLEAR, not guesses.

EXTRACT
- [field 1]
- [field 2]
- [field 3]

OUTPUT FORMAT
[table / bullets / one-pager structure]

DONE CHECK (print this checklist filled)
- [ ] Every number has a file pointer
- [ ] Every quote is exact
- [ ] Gaps listed under "Not in files"
- [ ] Interpretation separated from source claims

Worked toy example (sanitized)

Suppose you have three redacted files for a vendor decision. You are not pasting real customer data. You care about uptime SLA, support hours, and list price.

After an index pass, a good extract table might look like this (toy numbers for teaching):

Requirement	Vendor claim	Pointer	Status
Uptime SLA	99.9% monthly	vendor_a.pdf §3.2	Pass if we accept monthly window
Support hours	9 to 5 US Eastern weekdays	vendor_a.pdf §5.1	Fail vs 24/7 requirement in rfp.pdf §2
Price per seat	$48 / seat / month annual	pricing.csv row “Pro annual”	Pass under $50 cap
Data residency	Not stated	(none)	UNCLEAR / Not in files

Notice the last row. “Not stated” is a successful use of the tool. Inventing “US-only hosting” because that is common in the industry would be a failed use. Your job is to carry the UNCLEAR row into the human meeting, not to smooth it away for ChatGPT’s pride.

What good looks like: a short dialogue

Here is a compressed version of a clean session. Your real session will be longer. The shape is what matters.

You: Project “Vendor A decision.” Files: rfp.pdf, vendor_a.pdf, pricing.csv. Index pass only. No recommendations.

ChatGPT: Lists sections in both PDFs, columns in the CSV, flags one scanned appendix as unreadable, ends with “I have not analyzed content yet.”

You: Extract uptime SLA, support hours, list price per seat, data residency. Table with requirement, claim, pointer, status. No invented fields. Mark missing as UNCLEAR.

ChatGPT: Fills three rows with pointers. Data residency UNCLEAR. You open vendor_a.pdf §3.2 and confirm the 99.9% line before you paste the table into the decision doc.

That is the whole job in miniature: inventory, constrained extract, human verify. No heroic single prompt. No secret dump. No fake confidence on the blank field.

File types and awkward formats

Not every attachment behaves the same. Adjust your expectations:

Format	Usually fine for	Watch out for
Text PDF / DOCX	Section maps, quotes, policy extract	Headers/footers mistaken for body; multi-column layouts
Scanned PDF	Rough orientation if OCR works	Garbled text; verify every number by eye
CSV / XLSX	Column inventory, filters, simple comparisons	Wrong sheet; date formats; silent unit mixups
Slide decks	Claim lists, “what the deck promises”	Speaker notes missing; charts without data tables
Images of tables	Quick read when text export is impossible	Misread digits; always retype critical numbers from the source system
Mixed zip of everything	Almost never a good first move	Start with three named files max, then add

If a chart in a deck matters, ask for the underlying number from the table or the finance system, not from a screenshot description. Models misread chart pixels more often than people expect. When the number is load-bearing, go to the system of record.

When Chat is enough vs when to change doors

This series stays in everyday Chat on purpose. Still, multi-file work is where people start wondering about Work mode (agentic office steps) or Codex (repo work). Use the product map when the job changes shape:

Stay in Chat for read, extract, compare, rewrite, and brief writing against a small set of attachments you control.
Consider Work when the product needs multi-step office deliverables with apps, longer agent runs, and approval loops (covered in the later Work tutorial).
Consider Codex when the files are a software repo and the output is code changes, not a management brief.
Consider a Custom GPT only after the recipe is stable and weekly, not while you are still inventing the process.

If you do not know which door you are standing in, open ChatGPT product map before you grant broader permissions. More power is not the same as more truth.

Common mistakes

Mistake	What it looks like	Fix
One-shot the novel	“Summarize this 80-page PDF”	Chunk + index first
No file labels	Claims with no pointer	Require file + section on every fact
Secret soup	Full CRM export “just this once”	Redact or use approved workspace only
Stale Project files	v2 and v5 both uploaded	Delete superseded files; date names
Trusting quotes	Pretty quotation marks, never checked	Quote hunt before send
Wrong grain	Company-wide story from one region CSV	State grain in the goal and extract list
Essay without decision	Beautiful summary, no recommendation structure	Start from the decision in Step 1
Agent leap too early	Jump to Work/Codex for a simple extract	Finish the Chat recipe; change doors only when the job needs it

Practice this week

Pick one long public PDF (a product terms page, an open report, or a sanitized internal doc you are allowed to use).
Run an index pass. Save the section list.
Ask three specific questions with citation rules. Verify every quote by hand.
If you have a second file (even a tiny CSV of toy numbers), run the multi-file recipe end to end.
Optional: create one Project for a real recurring binder. Add instructions that ban invented metrics.

Part 4 shifts from documents to learning: explaining hard topics simply, quizzes, teach-back, and study habits that beat copy-paste theater. Part 5 will cover email, meetings, and workplace writing.

Quick recap

Long docs fail when you skip structure: use chunk → index → verify.
Demand file pointers and exact quotes for numbers that matter.
Use a Project when the same materials will stick around; still delete stale files.
Run the multi-file recipe: goal, files list, extract list, format, done check.
No secrets in consumer tools; prefer redacted samples and policy-approved workspaces.
Stay in Chat for extract and brief work; open the product map when the job needs Work or Codex.

Sources

Research and further reading used for this article. Product limits and UI labels change; verify on live pages before policy or training materials.

OpenAI Help: Projects in ChatGPT (workspaces that group chats, files, and instructions)
OpenAI Help: File uploads FAQ (size limits, rate and storage notes; confirm current numbers)
OpenAI Help: File storage and Library in ChatGPT (Library on web, storage separate from daily attachment limits)
OpenAI Help: Data controls FAQ (training and control orientation; pair with workplace policy)
OpenAI Help Center (live product help for ChatGPT features)
ChatGPT (product entry)
OpenAI policies (terms and usage policies)
Analytics Made Simple: Learn ChatGPT from scratch (foundation series)
Analytics Made Simple: ChatGPT product map (Chat vs Work vs Codex and related surfaces)
Analytics Made Simple: ChatGPT everyday tutorial (this series home)
Analytics Made Simple: Learn (paths hub)