PDF to Summary

---
name: pdf-to-summary
description: Turns a long PDF, report, contract or slide deck into a layered summary (three bullets, then one page, then full notes where every point cites its page) plus a watch list of dates, costs and obligations. Use when a document is too long to read in full but too important to skim blindly.
---

# PDF to Summary

You are the person who has already read the whole document, so the user does not have to be. Your job is to make one document readable at three depths without the three depths ever disagreeing with each other.

## Workflow

1. **Ask what the summary is for.** Decision, review, or general understanding. This changes what you keep far more than the document's subject does.
2. **Read the shape before the words.** Headings, section lengths, where the argument turns, where the numbers live.
3. **Write the core message in one sentence.** What the document is actually saying or asking for.
4. **Build the three layers top down.** Three bullets, then one page, then full notes with a page reference on every point.
5. **Sweep for watch items.** Dates, money, obligations, and anything that happens automatically.
6. **Name the gaps.** What the document leaves unclear, and what you could not read.

Copy this checklist into your working notes and tick it as you go:

```text
Document summary checklist
[ ] Purpose known: decide / review / understand, and who reads the summary
[ ] Structure mapped before any summarising
[ ] Core message written in one sentence
[ ] Layer 1: three bullets
[ ] Layer 2: one page, organised by section
[ ] Layer 3: full notes, every point carries a page or section reference
[ ] Layers checked against each other - no contradictions
[ ] Watch items listed: dates, money, obligations, automatic events
[ ] Gaps listed: unclear, missing, or unreadable
[ ] Coverage stated: which pages were actually read
```

## Step 1: what the summary is for

Ask for the purpose in the same message as the document. One question, four parts: what you need this for, who reads the summary, which sections matter most, and whether this is the whole document or an extract.

| The user needs to | Keep hard | Let go |
|---|---|---|
| Decide something | Options, costs, deadlines, what happens if they do nothing | Background, methodology, history |
| Review or approve | Changes from the last version, risks, anything they would be signing up for | Sections that are unchanged boilerplate |
| Understand a subject | The argument, the evidence, the limits of the evidence | Commercial terms, process detail |

If the user does not answer, assume "understand", say that you assumed it, and offer to re-cut the summary for a decision if that is what they actually need.

## Step 2: read the shape before the words

Spend the first pass on structure only. You are looking for where the document keeps its real content, which is rarely where it announces it.

- Count the sections and note which ones are long. Length is where the effort went.
- Find the turn: the paragraph where the document stops describing and starts asking, recommending, or obliging.
- Find where numbers appear. Numbers concentrate in a few places; note the page.
- Note anything the document points at but does not include: an appendix, a schedule, a linked policy, a referenced study.

`references/document-types.md` lists where the real content hides in contracts, research reports, slide decks, policies and academic papers. Read it when the document type is one you have not summarised in this session.

## Step 3: the core message in one sentence

Before writing any layer, write the sentence the whole document exists to deliver:

- "This contract renews itself every 12 months unless we cancel 60 days out, and the price rises 7% each renewal."
- "The study found a small effect, and its own authors say the sample was too narrow to generalise."
- "The team is asking for two more engineers and a delayed launch date."

If you cannot write that sentence, the document may genuinely lack one. Say so in those words rather than manufacturing one.

## Step 4: the three layers

Write them in order. Each layer is a longer version of the same truth, never a different one.

**In 3 bullets.** For someone who reads only this. What it is, the one number or date that matters most, and the thing they have to do or decide. No sub-bullets, no page references, no hedging.

**In one page.** Organised by the document's own sections, in plain sentences, roughly 250 to 400 words. Follow the document's structure, not your own, so the reader can navigate back into it.

**Full notes.** Every substantive point, each ending with a reference like `[p. 14]` or `[Section 4.2]`. Group by section. This layer is for the person who will act on it and needs to check it.

Then read the three layers back to back. If the one-pager says the contract is annual and the notes say it is 18 months, the notes are usually right and the shorter layer needs fixing. Never ship layers that disagree.

## Step 5: watch items

A separate list, after the layers, because these are what people miss and later regret. Sweep for:

- **Dates:** deadlines, notice periods, expiry, review dates, anything with a clock on it.
- **Money:** totals, per-unit costs, increases, penalties, minimums, what is excluded.
- **Obligations:** what the reader's side must do, by when, and what happens if they do not.
- **Automatic events:** auto-renewal, automatic escalation, default settings that apply if nobody acts.
- **Limits:** liability caps, exclusions, conditions that void something.

Each watch item gets one line and a page reference. Say plainly if the sweep found nothing in a category rather than silently omitting it.

## Step 6: gaps and coverage

Close with two short lists:

- **Unclear or missing:** terms used but never defined, a promise with no date attached, a number with no basis, a referenced appendix that was not included.
- **Coverage:** which pages you actually read. If the user pasted 12 pages of a 40-page document, say that in the first line of the summary, not in a footnote.

## Citing pages when there are no page numbers

Pasted text and web-exported documents often have no page numbers. Cite anyway, using whatever anchor the reader can search for:

- Numbered sections or clauses: `[Section 4.2]`, `[Clause 11(b)]`.
- Headings: `[under "Termination"]`.
- Nothing else available: quote the first four words of the sentence, `["The Supplier shall provide..."]`.

A note the reader cannot verify is worth much less than one they can.

## Keep the document's claim separate from your read

Write what the document says. When you add your own inference, label it in the same sentence.

- Document says: "Uptime was 99.2% across the quarter [p. 8]."
- Your read: "Reads as a miss - the contract commits to 99.5% [p. 3], though the document never connects the two."

Never merge the two. A reader who cannot tell your inference from the source will either trust too much or trust nothing.

## Example

**Input:** a 9-page renewal quote and terms from a cloud storage vendor. Purpose: decide whether to renew. Reader: the office manager who will sign it.

**Core message:** "Renewing as offered costs 41% more than last year, and the new terms lock in for 24 months instead of 12."

**In 3 bullets**

- Renewal is 24 months at $1,410/month, up from $1,000/month on a 12-month term.
- The price rise is mostly a storage tier change, not a rate rise: the actual usage sits just above a tier boundary.
- To avoid it, either cancel by 30 April or ask to be moved back to a 12-month term before signing.

**In one page**

> Pricing and term (pp. 1-3). The quote covers 24 months from 1 June at $1,410 per month, billed annually in advance. The previous agreement was 12 months at $1,000 per month, billed monthly...

**Full notes (extract)**

- Monthly fee $1,410, annual prepayment, no monthly option offered [p. 1].
- Term is 24 months with no break clause [p. 2].
- Storage tier moves from 5TB to 10TB; current usage is 5.3TB [p. 4].
- Cancellation requires 30 days written notice before the current term ends, so by 30 April [p. 7].

**Watch items**

- 30 April cancellation deadline: after this the current term auto-renews [p. 7].
- Annual prepayment means the full $16,920 leaves in one payment [p. 1].
- Overage above 10TB is charged at $0.09/GB, uncapped [p. 5].

**Unclear or missing:** the quote references "Schedule B - Support Levels", which was not included [p. 6].

**Coverage:** all 9 pages of the quote. The linked master agreement was not provided.

Notice that the three bullets are enough to act on, the notes are enough to argue with, and every number can be found again.

## Rules

- Do not add information that is not in the document. If the user needs outside context, say what is missing and offer to look it up as a separate step.
- Never let a shorter layer contradict a longer one.
- Every point in the full notes carries a reference. No exceptions.
- Flag legal, financial, medical and safety items in the watch list. Never bury them inside a paragraph.
- Use the document's own terms, defined once, rather than substituting friendlier words that shift the meaning.
- Say what you could not read: scanned pages, cut-off tables, images with text in them.
- Summarising a contract is not legal advice, and summarising a diagnosis is not medical advice. Say so once, plainly, and keep going.

## Edge cases

- **Only part of the document was provided.** Summarise what you have and put the coverage line first, not last. Do not extrapolate the missing sections.
- **The document contradicts itself.** Report both versions with both page references and say which one governs if the document says so. Do not silently pick one.
- **The text is garbled or scanned badly.** Summarise the readable parts, list the pages that came through as noise, and ask for those pages again rather than guessing at the numbers.
- **Almost all of it is boilerplate.** Say so. Three bullets, a short watch list, and one line naming which clauses are non-standard is a better summary than a padded one.
- **The user wants the document to support a decision it does not support.** Give the honest read, in one sentence, before the layers.
- **Numbers appear only in charts or images.** Say which numbers came from a chart and that they are read off, not stated.
- **Several documents at once.** Summarise each separately first, then add a short section on where they disagree. Merging them straight away hides the disagreement, which is usually the interesting part.
- **The document is a slide deck.** The speaker carried half the meaning and is not in the file. Summarise what the slides claim and flag where a slide clearly needed narration.

## Files

- `references/examples.md` - three documents worked end to end: a vendor contract read for a renewal decision, an industry research report read for a team briefing, and an internal strategy deck read for what it avoids saying. Read one fully before your first summary to see how far the layers travel apart.
- `references/document-types.md` - where the real content hides by document type, and the watch items each type reliably produces. Read it when you are handed a type you have not summarised yet.

# Worked examples

Three documents taken from the raw ask through to the finished layered summary. Read one all the way down before summarising your own.

## Contents

- [Example 1: a vendor contract read for a renewal decision](#example-1-a-vendor-contract-read-for-a-renewal-decision)
- [Example 2: an industry report read for a team briefing](#example-2-an-industry-report-read-for-a-team-briefing)
- [Example 3: an internal strategy deck read for what it avoids](#example-3-an-internal-strategy-deck-read-for-what-it-avoids)
- [What the three have in common](#what-the-three-have-in-common)

---

## Example 1: a vendor contract read for a renewal decision

### The ask

"We have 34 pages of terms from our logistics provider and the renewal is in six weeks. I need to brief our operations director. She will read three lines and then ask me questions."

Purpose: decide. Reader: an operations director who will decide, not read.

### Core message

"The renewal locks in three years at a rate that rises with fuel but never falls with it, and the exit clause costs six months of fees."

### In 3 bullets

- Three-year term from 1 September, minimum spend 4,000 pallets a year; last year's actual volume was 3,100.
- The fuel surcharge moves up with the index but has a floor, so the rate cannot come back down below the September figure.
- Leaving early costs six months of the minimum spend, roughly 62,000 at current rates.

### In one page

> **Term and volume (pp. 2-5).** Three years from 1 September, replacing the current one-year rolling arrangement. The agreement sets a minimum annual volume of 4,000 pallets. Volume below the minimum is charged as if the minimum were met, so the shortfall is billed rather than forgiven. Recorded volume for the twelve months to July was 3,100 pallets.
>
> **Rates and surcharges (pp. 6-11).** Base rates are fixed for the first year and then reviewed annually against the published haulage index. A fuel surcharge applies on top, recalculated monthly. The surcharge tracks the index upwards without limit and downwards only to a floor set at the September rate.
>
> **Service levels (pp. 12-19).** On-time delivery target of 96%, measured monthly and averaged over each quarter. Missing the target produces a service credit of 2% of that month's invoice, capped at 6% a year. There is no right to terminate for repeated failure.
>
> **Exit and renewal (pp. 20-24).** Termination for convenience is available after month 12 with 90 days notice and an early termination charge of six months minimum spend. The agreement renews for a further twelve months unless cancelled 90 days before the end of the term.
>
> **The rest (pp. 25-34).** Insurance, liability, data handling and dispute resolution, all close to the provider's standard wording.

### Full notes (extract)

- Term: 36 months from 1 September, no break in year 1 [p. 2].
- Minimum annual volume 4,000 pallets; shortfall billed at the standard pallet rate [p. 4].
- Actual volume to July stated in the covering letter as 3,100 pallets [covering letter, p. 1].
- Base rate fixed for 12 months, then indexed annually, upward review only [p. 7].
- Fuel surcharge recalculated monthly, floor set at the commencement figure [p. 9].
- On-time target 96%, quarterly average, 2% monthly credit, 6% annual cap [pp. 13-14].
- No termination right for persistent service failure [p. 15, and absent from p. 21].
- Termination for convenience after month 12, 90 days notice, six months minimum spend payable [p. 21].
- Auto-renewal for 12 months unless cancelled 90 days before expiry [p. 22].

### Watch items

- **The volume floor is above current usage.** At 3,100 pallets against a 4,000 minimum, roughly 900 pallets a year would be paid for and not used [pp. 4, and covering letter].
- **The surcharge floor is one-way.** Fuel falling does not reduce the bill below the September figure [p. 9].
- **Exit costs six months of the minimum, not six months of actual spend** [p. 21].
- **Auto-renewal 90 days out.** Put 3 June in the calendar now [p. 22].
- **Service failure has no exit.** Credits are capped at 6% a year and cannot be escalated to termination [pp. 14-15].

### Unclear or missing

- "Standard pallet rate" is used on p. 4 to price the shortfall and is never defined in the body. Schedule 2 was not included.
- The covering letter mentions a "loyalty adjustment" that appears nowhere in the terms.

### Coverage

All 34 pages of the terms and the two-page covering letter. Schedule 2 (rate card) was referenced but not provided.

**Why this summary is shaped this way:** the director asked for three lines, so the three bullets each carry a number she can challenge. The watch list, not the one-pager, holds the two facts that would change the decision (the volume floor and the one-way surcharge) because those are the parts a reader would otherwise meet for the first time in an invoice.

---

## Example 2: an industry report read for a team briefing

### The ask

"A 62-page report on remote work productivity got sent round the leadership channel. I have to summarise it in Monday's team meeting. It is being used to argue for return-to-office."

Purpose: understand, with an edge of review, because the report is being used as evidence. Reader: a team, verbally.

### Core message

"The report finds a small productivity gap between remote and office workers, and its own methodology section says the comparison is not like for like."

### In 3 bullets

- The headline finding is a 4% productivity gap in favour of office-based workers, measured over one quarter.
- The two groups were not comparable: office-based roles in the sample skewed towards support functions with per-ticket measurement, remote roles towards project work measured by self-report.
- The report's own limitations section says the finding "should not be used to compare individual arrangements", which is close to the use it is being put to.

### In one page

> **What was measured (pp. 8-14).** A survey of 1,900 employees across 41 companies, plus productivity data supplied by 12 of those companies. Fieldwork ran October to December. Productivity was defined differently by different participating employers; the report normalises but does not reconcile these definitions.
>
> **The headline finding (pp. 15-22).** Office-based workers recorded 4% higher output on the normalised measure. The gap widens to 9% for employees in their first year and disappears entirely for employees with more than five years tenure.
>
> **What the report says about its own limits (pp. 51-54).** Self-selection in the employer sample, one quarter of data, no control for role type, and an explicit note that the finding is about populations rather than individuals.
>
> **Funding and authorship (p. 2 and p. 61).** Commissioned by a workplace real-estate association. Authored by an independent research firm.

### Full notes (extract)

- Sample: 1,900 employees, 41 companies, 12 supplying data [p. 9].
- Fieldwork October to December, one quarter only [p. 9].
- Productivity definitions vary by employer, normalised not reconciled [p. 11].
- Headline gap 4% in favour of office-based [p. 16].
- Gap by tenure: 9% under one year, 0% over five years [p. 19, Figure 7].
- Figure 7 footnote: tenure bands under one year contain 210 respondents [p. 19].
- Limitations: no role-type control [p. 52].
- Explicit statement that findings are population-level [p. 53].
- Commissioned by a workplace real-estate association [p. 2].

### Watch items

- The tenure finding is the more interesting one and it is not in the executive summary: the gap is an onboarding effect, not a location effect [p. 19].
- The commissioning body has a commercial interest in the conclusion [p. 2]. This does not make the research wrong; it belongs in the briefing.
- One quarter of data, in the quarter containing the holiday period [p. 9].

### Unclear or missing

- How the differing productivity definitions were normalised is described in one sentence and never shown.
- No breakdown of the 41 companies by industry, so the reader cannot judge whether the mix resembles their own.

### Coverage

All 62 pages. Figures read from charts where no table was given, noted in the full notes as read off.

**Why this summary is shaped this way:** the report is being used as an argument, so the methodology moved to the front and the limitations section became a headline rather than an appendix. The tenure split gets its own bullet because it is the finding that would actually change what the team does.

---

## Example 3: an internal strategy deck read for what it avoids

### The ask

"Our new product lead sent a 40-slide strategy deck ahead of Thursday. I need to know what she is actually asking for and what she has skipped."

Purpose: review. Reader: a peer preparing to discuss it.

### Core message

"The deck asks for four engineers and a six-month runway to rebuild the onboarding flow, and it does not show what happens to the existing roadmap if it gets them."

### In 3 bullets

- The ask is four engineers for six months, on slide 34 of 40.
- The case rests on one number: 62% of new accounts never complete setup. That number comes from a cohort of accounts created in one month.
- The deck does not show the cost side: no slide covers what the four engineers stop doing, and the existing roadmap is not mentioned after slide 6.

### In one page

> **The problem framing (slides 3-11).** Onboarding completion is 38%; competitor benchmarks are cited at 55-70%. The 62% drop-off figure is drawn from accounts created in March, described on slide 9 as "a representative month".
>
> **The diagnosis (slides 12-22).** Six friction points, four of which are supported by session recordings and two by team opinion. The strongest evidence sits on slides 16 and 18, where the drop-off is tied to a specific verification step.
>
> **The proposal (slides 23-33).** A staged rebuild, three phases, with the verification step redesigned first. Phases two and three are described in one slide each and have no estimates.
>
> **The ask (slides 34-36).** Four engineers, six months, with a check-in after phase one.
>
> **Appendix (slides 37-40).** The cohort data behind the 62% figure, and a sensitivity table that is not referenced anywhere in the body.

### Full notes (extract)

- Current completion rate 38% [slide 4].
- Competitor benchmark range 55-70%, source cited as an industry blog post [slide 5, footnote].
- 62% drop-off from March-created accounts, n = 1,240 [slide 9 and appendix slide 38].
- March described as "representative"; no comparison months shown [slide 9].
- Six friction points; evidence for points 5 and 6 is described as "team observation" [slides 20-21].
- Verification step accounts for 41% of the drop-off [slide 18].
- Phase one estimated at 10 weeks; phases two and three unestimated [slides 27, 30, 32].
- Ask: four engineers, six months [slide 34].
- Appendix sensitivity table shows the payback case failing if completion improves by less than 9 points [slide 40, unreferenced].

### Watch items

- **The appendix contradicts the confidence of the body.** Slide 40 shows the case failing below a 9-point improvement, and the body never mentions a threshold [slide 40].
- **Two of six friction points have no evidence behind them** [slides 20-21].
- **Phases two and three carry no estimate** but are inside the six-month ask [slides 30, 32].
- The competitor benchmark is sourced from a blog post, not a study [slide 5].

### Unclear or missing

- What the four engineers would otherwise be building. The roadmap appears on slide 6 and never again.
- Whether March is representative. No other month is shown.
- Who owns the decision after the phase-one check-in.

### Coverage

All 40 slides including the appendix. This is a deck built to be presented; roughly a third of the slides are single images or numbers that clearly depend on narration, and those are marked in the full notes.

**Why this summary is shaped this way:** for a review, the absences matter as much as the claims, so "unclear or missing" carries real weight rather than being a formality. The appendix contradiction is the first watch item because it is the single fact most likely to change Thursday's conversation.

---

## What the three have in common

1. The purpose was settled before a word was summarised, and it visibly changed what got kept.
2. The core message is one sentence containing at least one number or date.
3. Every point in the full notes can be found again in under ten seconds.
4. The watch list holds what a reader would otherwise discover too late, not a repeat of the summary.
5. Where the summary adds a judgement, the judgement is labelled and the source is still visible underneath it.
6. Coverage is stated plainly, including what was referenced but never provided.

# Where the real content hides, by document type

Every document type buries its important parts somewhere predictable. Read the row that matches before your first pass.

## Contents

- [Contracts and terms](#contracts-and-terms)
- [Research and industry reports](#research-and-industry-reports)
- [Slide decks exported to PDF](#slide-decks-exported-to-pdf)
- [Policies, handbooks and procedures](#policies-handbooks-and-procedures)
- [Academic papers](#academic-papers)
- [Financial statements and annual reports](#financial-statements-and-annual-reports)
- [Quick reference](#quick-reference)

---

## Contracts and terms

**Read in this order:** the definitions, then the obligations on the reader's side, then termination, then money, then liability. The first pages are usually the least interesting.

Where the content hides:

- **Definitions.** A term defined loosely at the front changes the meaning of every clause that uses it. "Business day", "Services", "Confidential Information" and "Affiliate" are the usual suspects.
- **Schedules and annexes.** The price, the service levels and the scope often live here, not in the body. If a schedule is referenced and not attached, that is a gap, not a detail.
- **Cross-references.** "Subject to clause 14" means clause 14 is doing the work.
- **The one-way clauses.** Scan for obligations that bind only one side: notice periods, audit rights, termination for convenience.

Watch items this type reliably produces: auto-renewal, notice period, price escalation, minimum commitment, liability cap, exclusivity, what happens to data on exit.

---

## Research and industry reports

**Read in this order:** the methodology, then the findings, then the executive summary. Reading the executive summary first anchors you to the author's framing.

Where the content hides:

- **Methodology and sample.** Sample size, who was surveyed, when, and who paid for it. A report on cloud spending sponsored by a cloud vendor is still useful, but the reader must be told.
- **Footnotes on charts.** "Excludes respondents who did not answer" and similar notes routinely change what a headline number means.
- **The limitations section.** Authors are often honest here and nowhere else.
- **The gap between the chart and the caption.** Read the axis, not the caption.

Watch items this type reliably produces: who funded it, sample size and date, definitions that differ from the industry norm, projections presented in the same style as measurements.

---

## Slide decks exported to PDF

**Read in this order:** the appendix, then the body, then the title slide. The appendix holds the numbers the body summarises.

Where the content hides:

- **The appendix.** Detailed tables, assumptions, sensitivities.
- **Speaker notes,** if the export kept them. Often the only place the caveats appear.
- **Axis labels and footnotes** on charts, which are usually set small deliberately.
- **The missing slide.** A deck arguing for a decision that never shows the cost of the alternative has told you something.

Watch items this type reliably produces: forecasts with no stated assumption, before-and-after comparisons with different time windows, a request buried in the second-to-last slide.

Always flag that a deck was built to be narrated. Half the argument was spoken.

---

## Policies, handbooks and procedures

**Read in this order:** scope (who it applies to), then obligations, then exceptions, then the process for raising something.

Where the content hides:

- **Scope and applicability.** A policy that applies only to permanent staff is a different document to one that covers contractors.
- **The exceptions list.** This is where the real rule lives.
- **Version and effective date.** Superseded policies circulate for years.
- **The escalation path.** Who decides, and how long they have.

Watch items this type reliably produces: effective dates, mandatory reporting timeframes, disciplinary consequences, anything requiring written approval.

---

## Academic papers

**Read in this order:** abstract, then figures, then limitations, then conclusion. The methods section only if the finding matters to the user's decision.

Where the content hides:

- **The limitations paragraph** near the end of the discussion.
- **Effect size versus statistical significance.** A significant result can still be tiny. Report both when the paper gives both.
- **Pre-registration and conflict-of-interest statements.**
- **What the paper actually measured** as opposed to what the title implies it measured.

Watch items this type reliably produces: sample size, whether it was a controlled study or an observational one, funding source, whether the conclusion is stated more strongly than the results support.

---

## Financial statements and annual reports

**Read in this order:** the notes to the accounts, then the cash flow statement, then the narrative sections.

Where the content hides:

- **Notes to the accounts.** Accounting policy changes, one-off items, related-party transactions.
- **Cash flow versus profit.** Profit is an opinion; cash is a fact. When the two diverge, that divergence is the story.
- **Restated prior-year figures.** If last year's numbers changed, find out why.
- **Adjusted metrics.** Anything labelled "adjusted", "underlying" or "normalised" has had something removed. Find out what.

Watch items this type reliably produces: covenant conditions, going-concern language, contingent liabilities, changes in accounting treatment.

---

## Quick reference

| Type | Read first | Most common trap |
|---|---|---|
| Contract | Definitions, then obligations | Terms hidden in an unattached schedule |
| Research report | Methodology | Executive summary framing the reader before the evidence does |
| Slide deck | Appendix | Missing narration and unlabelled axes |
| Policy | Scope and exceptions | Reading a superseded version |
| Academic paper | Abstract, figures, limitations | Conclusion stated more strongly than results support |
| Financial report | Notes to the accounts | Adjusted figures presented as the headline |

When a document is a mix of several types, treat each part with its own rules and say in the summary which parts you treated which way.