Free to read · no account needed

16 min · research

One clean pass over a long document with Claude

Goal: extract decisions, risks, and quotes from a long PDF or thread—without inventing content.

Best for: Claude (Projects help when you revisit the same pack of files). Strong fit for reports, RFPs, papers, long email exports.

Long docs need section-scoped asks. Freeze facts; polish tone in a later pass.

Why beginners skip this

Skipping the hard part of “One clean pass over a long document with Claude” feels fine until a live thread pays the bill. Hold: Chunk + running decision ledger; never treat a mega-summary as the source of truth. Mid-week cost shows up as rework, not as a missing bookmark.

Prep the input

1. Prefer oldest → newest for threads; strip duplicate signatures. 2. Tell Claude what “done” is: timeline · decisions · open questions · quote sheet—one shape per prompt. 3. For very long material, chunk → mini-summaries → reconcile (see below).

Prompt — decision ledger

Goal: Build a decision ledger from the document only.
Format: table Date | Decision | Who | Evidence quote (≤25 words) | Status (decided/proposed/unclear)
Boundaries:
- If not stated, write “not in source”.
- Do not infer agreement from polite language.
- Number rows. Flag contradictions in a short note under the table.
Document:
---
[paste or attach]
---

Chunk-and-merge (50+ pages / huge threads)

Example

Filled prompt

Goal: Build a decision ledger from the Q3 planning PDF only.
Format: Date | Decision | Who | Evidence quote (≤25 words) | Status
Boundaries: If not stated, “not in source”. Do not infer agreement from polite language.

What good output looks like

  • Table rows only for decisions actually present in the PDF chunk
  • Polite “sounds good” left out of Decided
  • Contradictions listed under the table—not smoothed away
  • Each Decided row has a cite hint (page/section) you can verify
  • No invented budget numbers, owners, or meeting dates
This is Chunk [n] of [N]. Summarize: running decisions, open questions, names/dates to watch.
Max 12 bullets. Quote anything that looks like a commitment.

Then:

Reconcile these chunk summaries into one ledger.
Merge duplicates. Mark items raised in early chunks and closed later as Settled.
List residual open questions only.

Spot-check rule

Example

Filled prompt

This is Chunk 2 of 4. Summarize: running decisions, open questions, names/dates to watch. Max 12 bullets.
Then reconcile all chunk summaries into one ledger; mark Settled items; list residual open questions only.

What good output looks like

  • Chunk summary stays ≤12 bullets—no essay re-hash of the PDF
  • Merged ledger has no duplicate decision rows
  • Only truly open questions remain in Open
  • Names/dates to watch flagged for the next chunk
  • You can hand the ledger to a teammate without re-reading noise

Before you act: verify dates, amounts, names, commitments in the original. Models can shift a deadline by a day or treat a proposal as agreed.

For rows 3, 7, and 12: return the exact source sentence for each Evidence quote. If missing, mark the row Unreliable.

Decided vs polite noise

“Sounds good” and “let’s revisit” are not decisions. Your ledger should only promote rows with an owner, a clear choice, and a place you can cite in the original. Mark shaky rows Unreliable and do not act on them.

Example

Filled prompt

PDF section on headcount.
Decision ledger with quotes. "Sounds good" ≠ decided.
Context: real workstream this week (Asia/Shanghai). Name actual files/people.
Constraints: no invented metrics/owners/prices; label Unknowns; stop before send/delete/pay.
Done check: I can mark pass/fail in one minute on the output.

What good output looks like

  • Pulls from Project/file named in the ask for “extract decisions, risks, and quotes from a long P”
  • Standing instructions respected; no silent scope creep
  • Decisions quoted or labeled Unclear — never “sounds good”
  • One-minute check: does output match claude-long-document-pass Done?

Red flag

Mixing structure and tone in one pass; feedback with nothing to tick; “overall fine” as a review—redo single-layer.

Save this card

Card: Chunk + running decision ledger; never treat a mega-summary as the source of truth. Keep this beside the tool; read before you start.

1. What I shipped: ___ 2. What I refused: ___ 3. What I will never-do again: ___ (date: Asia/Shanghai)

  • [ ] Tools/writes paused
  • [ ] Refuse-line dated
  • [ ] Artifact path saved
  • [ ] Next link chosen (or stop)

Before you start

Open the Claude Project you will keep. Confirm standing instructions + the knowledge file for “One clean pass over a long document” are current. Guide: claude-long-document-pass. If you only have a blank practice box, stop and fetch the real job first.

  • Goal in one line: ___
  • Allowed tools: ___
  • Forbidden: ___
  • Human approve needed for: ___

How to use this week

This week in Claude: chunk one long PDF/thread and maintain a rolling decision table—never treat a single mega-summary as source of truth on a live artifact—not a blank demo. Write one refuse-line and date the note (Asia/Shanghai). If you only skimmed claude-long-document-pass, put 25 minutes on the calendar and redo once.

Pitfalls

Example

Filled prompt

Q3 planning PDF section on headcount.
Decision ledger: date|decision|who|quote≤25 chars|status. "Sounds good" ≠ decided.
List contradictions under table.

What good output looks like

  • Returns exact sentences for the row—or marks Unreliable
  • You spot-check dates/amounts in the original before acting
  • No action recommended on Unreliable rows
  • Cites chunk/page so you can jump back in the PDF
  • Refuses to “average” two conflicting figures into one fake number
  • Asking for summary + strategy + email in one go → split outputs
  • Letting Claude “fill gaps” → forbid inference in Boundaries
  • Skipping Projects for recurring packs → recreate context every time

Common pitfalls

PitfallFix
Repasting FAQ every chatPut standing instructions + one knowledge file in Project
“Summarize the PDF” with no ledgerDecision ledger: quote or Unclear
Deep thinking on every polishSave deeper mode for real tradeoffs

Practice now

Live card: one Claude artifact/diff/email pass for “One clean pass over a long document with Claude” with Blind before share (Asia/Shanghai).

Do a single-layer pass on a long doc (structure or facts only) and return a checkable edit list. Tick three edits today (Asia/Shanghai).

Today card (copy) — claude-long-document-pass: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

  • [ ] Artifact path written
  • [ ] Refuse example would say "yes"
  • [ ] Dated Asia/Shanghai

One sentence takeaway

Chunk + running decision ledger; never treat a mega-summary as the source of truth. This week: one clean single-layer pass with a dated ledger.

  • You can point to a dated refuse-line
  • You have one artifact path from tonight
  • You can teach the red flag in one sentence

Next

Email polish · Research path · Artifacts for deliverables · Extended thinking · Outline before prose

Pair with today: open the first Next link (claude-email-from-correct-to-polished) only after you dated a refuse-line for this guide (Asia/Shanghai).

← Back to home · Learn · Tools · Compare · Start