Reference
Frameworks: the intellectual foundation of document synthesis
Part of the Document Synthesis skill · loaded on demand from SKILL.md
Read this file when you need the reasoning behind a step in SKILL.md, when a grouping will not resolve, or when you have to defend a summary line to a reviewer who knows Minto.
Contents
- Minto's Pyramid Principle: what it claims and why it works
- The three rules for any grouping
- The four logical orders
- Inductive versus deductive grouping
- Vertical logic: the question-answer dialogue
- Summarising upward: the intellectually blank assertion
- MECE as a test, not a slogan
- The SCQ introduction
- The McKinsey "so what" discipline
- Summary versus synthesis
- Source-traceable extraction and reliability grading
- Contradiction handling: a taxonomy
- Why extraction precedes grouping: atomic claims
- Evidence weight and the writing of caveats
- Why the method holds at scale, and where it breaks
1. Minto's Pyramid Principle: what it claims and why it works
Barbara Minto developed the Pyramid Principle while at McKinsey in the 1960s and 1970s and published it as "The Pyramid Principle: Logic in Writing and Thinking" (Pitman, 1987), with a third edition, "The Minto Pyramid Principle: Logic in Writing, Thinking and Problem Solving" (Pearson, 2009). It is the basis of how strategy firms structure documents and, less obviously, how they structure the thinking that precedes them.
The claim is that the mind can hold only a small number of ideas at once and copes by grouping and summarising. A reader given a sequence of facts will impose a structure on them whether or not the writer intended one. Minto's illustration is a list of shopping items that becomes memorable only once it is sorted into categories. The consequence for writing is that the writer must do the grouping and summarising first, and present the summary before the detail, so that the reader's mind receives ideas in the order it would otherwise have to construct them.
For document synthesis the claim runs the other way as well. The documents in a corpus were each written with their own pyramid, answering their author's question. The synthesist's job is to dismantle those pyramids into claims and rebuild one pyramid that answers the client's question. Reading a corpus without extracting claims first means adopting the strongest author's pyramid by default, which is why prior consultant decks are read last.
2. The three rules for any grouping
Minto states three rules that any group of ideas in a pyramid must satisfy. They apply at every level, from the governing thought and its key lines down to the last group of claims under a sub-point.
Rule 1: ideas at any level must summarise the ideas grouped below them. The idea above a group is not a label for the group; it is the conclusion the group supports. If the group below were removed, the idea above should still make an assertion the reader could agree or disagree with.
Rule 2: ideas in each group must be of the same kind. All causes, or all effects, or all steps, or all recommendations. A group that mixes a cause with a recommendation cannot be summarised, because there is no single kind of statement that covers both. In extraction this rule appears as the type field in the ledger (fact, estimate, forecast, opinion, recommendation): claims of different types are rarely of the same kind.
Rule 3: ideas in each group must be in a logical order. Order is not decoration. The order tells the reader how the ideas relate, and an order that cannot be named usually signals that the group is not a group. The four orders are set out in section 3.
The rules interact. A group whose members are not of the same kind has no natural order and cannot be summarised; a group that can be summarised only by an intellectually blank assertion (section 6) is usually not of the same kind. In practice, apply the rules in the order same kind, order, summary: fix membership first, then sequence, then write the line.
3. The four logical orders
Minto identifies four ways a group can be ordered, each reflecting a different mental process.
| Order | Mental process | When it applies | Test |
|---|---|---|---|
| Chronological (time) | Determining cause and effect, or steps in a process | Actions, steps, events | Would reversing two members change the meaning? |
| Structural | Dividing a whole into parts | Segments, regions, functions, components of a bridge | Do the parts add to the whole with no overlap? |
| Comparative (degree) | Ranking by importance or size | Reasons, drivers, options | Is the ranking criterion stated and applied consistently? |
| Deductive | Reasoning from premises to a conclusion | Arguments where one point depends on another | Does each step follow from the previous with "therefore"? |
For synthesis the two most used orders are structural (the margin bridge in the worked example is a structural decomposition of 180 bps into four parts) and comparative (key lines ordered by size of effect or by strength of evidence). Chronological order appears when the corpus spans years and the question is about what changed. Deductive order is reserved for the case in section 4.
When the order of a group cannot be named, the group needs work. Either it is two groups (different kinds mixed), or the summary line is blank (section 6), or one member does not belong.
4. Inductive versus deductive grouping
This is the distinction that most often goes wrong in synthesis, and Minto is precise about it.
Deductive grouping is a chain. The first idea is a premise about the world, the second is a related premise, the third is the conclusion that follows from the two together ("therefore"). Every step depends on the previous one. Minto's example form: all men are mortal; Socrates is a man; therefore Socrates is mortal. The summary of a deductive group is the conclusion, usually prefaced by the reasoning. Deductive groups are fragile: refute one premise and the conclusion falls.
Inductive grouping is a set. The ideas are of the same kind and independent of one another; together they support a conclusion but none depends on another. The summary states what the set has in common and what that implies. Inductive groups are resilient: remove one member and the conclusion stands, weakened.
Minto's guidance is to prefer inductive structure for the key line of a document, because readers find it easier to hold a set of parallel points than to follow a chain, and because a reader who disagrees with one premise of a chain stops reading. Deductive structure is reserved for the case where the argument genuinely is a chain and presenting it as a set would misrepresent the logic.
For synthesis the practical rule is: build the pyramid inductively unless the answer depends on a conditional. "The plan assumes list-price increases; customers already perceive the client as expensive; therefore the list-price component is at risk" is deductive and must be presented as such, because the third line is not a separate finding. The four key lines of the worked example are inductive: each stands on its own evidence, and the governing thought would survive the loss of any one of them.
Do not mix the two at one level. A group of three inductive lines with a fourth line that is a "therefore" from the first three is a two-level structure pretending to be one.
5. Vertical logic: the question-answer dialogue
The pyramid's vertical relationship is a dialogue. A statement at any level raises a question in the reader's mind; the level below answers it. The governing thought "the plan overstates recoverable margin" raises "why do you say that?"; the key lines answer it. Each key line raises "how do you know?" or "how much?"; the claims below answer that.
The vertical test for any line is to write the question the reader will ask on reading it, and check that the children answer that question and no other. Common failures: children that answer "what else is true?" (a list, not an answer), and children that answer a question the parent did not raise (usually a sign that the parent is blank).
For synthesis this test is the fastest way to find a key line that is really a topic. If the question the reader asks is "and?", the line has not asserted anything.
6. Summarising upward: the intellectually blank assertion
Minto's most practical rule for synthesis concerns the summary line above a group. She warns against what she calls intellectually blank assertions: statements such as "there are three reasons" or "we recommend five changes". These do not summarise the ideas below; they state the kind of idea that will follow. The reader learns nothing from the line and must read the whole group to find out what it says, which defeats the pyramid.
The rule for writing a non-blank summary depends on the kind of idea in the group.
Action ideas (the group tells the reader to do things): the summary states the effect of carrying out the actions. Not "four steps to recover margin" but "these four steps recover $3.0M within twelve months at no volume risk".
Situation ideas (the group tells the reader about the world): the summary states what the similarity among the ideas implies. Not "three findings on freight" but "freight is under-recovered because surcharge tables have not been updated since FY23".
Minto's method for finding the summary is to ask why these ideas were grouped together. The answer is the similarity; the similarity plus the implication is the summary line. The similarity usually lies in one of three places: the ideas share the same kind of subject, the same kind of predicate (action or object), or the same kind of judgement. When no similarity can be found, the group is not a group.
In the ledger this rule shows up as a discipline: a candidate group whose summary line cannot be written without naming a category gets split or merged before it reaches the pyramid.
7. MECE as a test, not a slogan
Minto coined MECE (mutually exclusive, collectively exhaustive) at McKinsey as the criterion for a sound grouping: the members do not overlap, and together they cover the whole of what the parent asserts. It is Rule 2 and Rule 3 combined, with a completeness check added.
MECE is applied as two questions about a group, in this order:
- Exclusive: is there a claim in the ledger that could belong to two of these groups? If yes, the boundary between the groups is undefined, and the summary lines will double-count. Fix by defining the boundary (by cause, by period, by segment) and reassigning claims.
- Exhaustive: is there a category of claim in the ledger, or an obvious category in the world, that none of the groups covers? If yes, either add the group or state in the summary that the perimeter excludes it.
The standard error is treating MECE as a property of a framework rather than of a grouping against a question. "Price, volume, mix, cost" is MECE for a margin bridge and useless for a question about customer churn. The test is always: MECE with respect to what the parent line asserts.
A note on tolerance. Perfect exhaustiveness is rarely available from a corpus. A residual (the 2 bps in the worked example) is acceptable and should be shown, because showing it is what proves the decomposition was attempted.
8. The SCQ introduction
Minto's introduction structure is Situation, Complication, Question, with the Answer (the governing thought) following. The situation is what the reader already knows and agrees with; the complication is what changed or went wrong that makes the question urgent; the question is what the reader now wants answered. The purpose is to establish relevance before asserting anything contentious: a reader who has agreed with the situation and recognised the complication will accept the question as theirs and read the answer as a response rather than an intrusion.
In a one-page synthesis the SCQ is two or three sentences. It is written last, after the governing thought exists, because the complication must be the one that the governing thought resolves. The problem-framing skill produces the SCQ for the engagement; the synthesis borrows it and narrows it to the corpus.
9. The McKinsey "so what" discipline
Ethan Rasiel described the working practices of McKinsey consultants in "The McKinsey Way" (1999) and, with Paul Friga, in "The McKinsey Mind" (2001). Among the practices described are fact-based, hypothesis-driven problem solving, "don't boil the ocean" (analyse what bears on the answer, not everything), the 80/20 rule, and the elevator test (state the answer in thirty seconds). Running through all of them is the habit of asking "so what?" of every chart, fact, and finding until the answer is something the client can act on.
The mechanism is that description and decision are different products. Documents in a corpus are descriptive: a board pack reports what happened, a consultant deck reports what was analysed, a survey reports what customers said. A decision-maker needs implications and actions. The so-what chain is the conversion:
- Finding: what the sources say, cited by claim ID.
- Implication: what it means for this client's question, given the other findings.
- Action: what the decision-maker should do, specifically enough to assign an owner.
The chain is run on every key line of the pyramid and written in three columns before the synthesis paragraph is drafted. Two tests apply. The first is that the action must be a verb with an object and a time or threshold; "consider pricing" fails, "remove the list-price component from the FY27 plan until a competitor benchmark exists" passes. The second is to ask "so what?" once more of the implication; if a further action appears, the first was an intermediate step.
The discipline prevents the fact dump, which is the most common output of a synthesis done under time pressure: accurate, well-cited, and of no use to the person who has to decide.
10. Summary versus synthesis
The distinction matters because the two look alike on the page and are different in value.
| Summary | Synthesis | |
|---|---|---|
| Unit of work | One document | A corpus |
| Structure | The source's | The question's |
| Claims | Subset of the source's claims | Claims from many sources, plus new claims that no source contains |
| What it reveals | The source's emphasis | Agreement, contradiction, and gaps across sources |
| Failure mode | Inherits the source's blind spots | Fabricates agreement where sources differ in definition |
| Role in this skill | Intermediate (running notes) | Deliverable |
The mechanism by which synthesis produces new claims is juxtaposition. Placing the consultant's mix estimate next to the finance memo's rebate reclassification, in a ledger where both carry dates and definitions, exposes that the first absorbed the second. Neither document says so; the synthesis does. This is also why extraction is done before grouping and grouping before writing: the new claims appear at the grouping stage, and a synthesist who starts writing from the documents never reaches it.
11. Source-traceable extraction and reliability grading
Traceability is a diligence discipline rather than a published framework, but it is universal in data-room work and in any engagement where a client will challenge numbers. The rule is that every assertion in the deliverable can be followed to a document and a page through a claim ID, and that the reliability of the source travels with the claim.
Reliability grading (A audited or system, B management-prepared, C third-party analysis, D opinion or undated) does three jobs. It sets reading order (A and B first). It decides genuine contradictions (higher grade wins, then recency, then proximity to the underlying system). It forces caveats into the synthesis text: a key line that rests on grade C must say so, because the reader will otherwise treat a consultant's estimate as a fact.
Grade the source, not the claim, and then note exceptions in the ledger. A consultant deck (C) that reproduces a table from the audited accounts (A) yields a claim whose note field says "reproduces D03 p.14; grade A by provenance". The reverse also happens: a board pack (B) that includes an unsourced market-share estimate yields a claim noted "unsourced; treat as C".
Claim types (fact, estimate, forecast, opinion, recommendation, gap) are separate from grades and equally important. A grade A document can contain a forecast; it is still a forecast.
12. Contradiction handling: a taxonomy
Two claims that conflict must be classified before they are resolved, because the classes have different resolutions and only one of them is a real disagreement.
| Class | What is happening | Resolution | Example |
|---|---|---|---|
| Definitional | Same name, different measure | State both definitions; pick the one that matches the question; convert the other | Gross margin before and after rebates |
| Temporal | Both true at different dates | Order by date; report the change as a finding | Headcount 412 in the plan, 396 in the latest pack |
| Scope | Different perimeters | Reconcile to the same perimeter or report separately | Group revenue versus continuing operations |
| Method | Different calculation, both defensible | Recompute from the most primary data available; document the method chosen | Mix effect at constant versus actual segment margins |
| Genuine | Incompatible facts or forecasts about the same thing at the same time | Grade, then recency, then proximity to system; if unresolved, park as open with the deciding evidence named | Two market-share figures with no method stated |
The order of classification is the order in the table, because the earlier classes are cheaper to check and more common. In practice most conflicts in a corporate corpus dissolve at definitional or scope; the ones that survive to genuine are the ones the client needs to hear about.
Two rules apply regardless of class. Never average: the mean of two numbers that measure different things measures nothing. Never resolve silently: every resolution is logged with an X-ID and reasoning, and the synthesis cites the X-ID rather than one of the original claims, so a reviewer can see that the conflict was seen.
13. Why extraction precedes grouping: atomic claims
Minto's rules cannot be applied to documents; they can only be applied to ideas. A board pack is not "of the same kind" as a finance memo, and neither can be "summarised upward" with another document. The unit that the rules operate on is the single assertion, which is why the ledger exists and why its rows are atomic.
Three properties of an atomic claim make the later stages work:
- It has one type. A row that mixes a fact with an attribution ("margin fell 180 bps because of freight") cannot be placed in a group of causes or a group of effects; split, the fact goes in the bridge and the attribution goes in the contradiction log where it belongs.
- It has one definition and one period. Two rows with the same metric name and different definitions are the raw material of a definitional contradiction; a row that hides two definitions in one number is the raw material of a wrong synthesis.
- It has one location. A claim that spans two pages or two documents cannot be checked in a minute. Where a document states the same thing twice, the note field carries the second location.
Atomic claims also make evidence counting honest. The strength of a key line is not the number of rows beneath it but the number of independent sources of grade A or B beneath it. Five rows from one board pack are one source. A consultant deck that reproduces the board pack's figure is not corroboration. The grouping worksheet in templates.md carries a column for independent A/B sources for this reason, and a key line with a count of one is written with a caveat.
14. Evidence weight and the writing of caveats
A synthesis that treats every claim as equally reliable will be wrong in a way the reader cannot detect. The remedy is to carry grade and type into the text, briefly and in the same sentence as the number:
| Situation | How the synthesis phrases it |
|---|---|
| Grade A fact | State it. "Gross margin fell to 22.8% (D03-C02)." |
| Grade B fact | State it; name the function if the reader might doubt it. "Sales operations reports 19% of quotes off-policy (D05-C03)." |
| Grade B or C estimate | Name it as an estimate and its basis. "Sales operations sizes the leakage at $1.2M, from one quarter of data (D05-C06)." |
| Grade C survey | State the sample and that it is stated intent where relevant. "27% say they would reduce spend, stated intent, n=412 (D06-C07)." |
| Grade D | Do not cite as evidence. Use it to point to a source that can be cited, or list it as a gap. |
| Forecast by an interested party | Name the party. "The consultant's plan assumes 65% realisation with no volume loss (D02-C13)." |
The caveat belongs in the sentence, not in a footnote, because executives read sentences. A key line that would collapse without a grade C claim is written so that the reader sees the dependency.
15. Why the method holds at scale, and where it breaks
The method holds at scale because none of its steps requires the whole corpus to be in view at once. Extraction is per document; grouping is over the ledger; contradiction handling is over pairs of claims; the pyramid is over groups. What breaks at scale is memory: the synthesist forgets what batch 2 found by batch 7, tags drift, and duplicate claims inflate the evidence count for whichever theme was extracted most often.
The remedies are procedural. The running notes hold the emerging pyramid so it does not live in working memory. The controlled tag vocabulary prevents drift. Cross-referencing duplicates (rather than logging them again) keeps the evidence count honest: five documents that all quote the same board figure are one source, not five. Consolidation every two batches catches drift early. Per-document claim numbering allows parallel extraction and painless merging.
Where the method breaks is at the boundary of the corpus. A synthesis can only say what the documents say and what follows from placing them side by side. When the gap list grows longer than the key lines, the corpus is insufficient and the work should hand off to boardroom-storm-research for external sources or primary-research-synthesis for interviews, with the gap list as the brief.