Full SKILL.md
Document Synthesis
Synthesises a set of supplied documents (data rooms, board packs, prior consultant decks, transcripts) into a source-traceable answer — graded document index, claim ledger with page pointers, contradiction log, Minto pyramid, one-page synthesis and gap list. Use whenever the user hands over more than three documents and wants one answer, one fact base or "what do we already know", or says "synthesise these", "what does the data room say", "key points from this board pack", "reconcile the consultants with finance", "these decks don't agree", "themes across these reports". Not for sources that must be found (boardroom-storm-research) or interview material (primary-research-synthesis).
Turns a corpus into atomic, cited claims, then into an argument. Method sources (Minto's grouping and pyramid rules, the so-what discipline) are in references/frameworks.md; read it only when a step needs defending.
When not to use
- Sources have to be found:
boardroom-storm-research. - Transcripts, expert calls, field notes:
primary-research-synthesis(codes speech rather than extracting claims). - The question is unclear:
problem-framingfirst; a synthesis without a question is a summary. - The job is the client deck or memo:
executive-communication. The file:docx,pptx,xlsx.
Inputs
Ask for, or infer and label: the question in one line and who acts on the answer (use a problem-framing statement verbatim if one exists); the corpus (file list, types, dates, authors, page count) and which documents the user trusts most and least; hypotheses in play; deadline and depth (one page by tomorrow means triage; a diligence fact base means full extraction); quoting constraints.
Ask one round of questions if the question or the corpus is missing. Otherwise write the assumed question at the top of the running notes and proceed.
Workflow
-
Question and perimeter. One-line question, two or three sub-questions (a small issue tree; more than five branches goes back to
problem-framing), perimeter (entity, period, metrics), output form. Done when the question alone tells you which claims are irrelevant. -
Inventory and triage. Document index (
references/templates.md§2): ID D01…, title, author, date, type, pages, reliability grade, read depth. Grade the source not the claim: A audited, filed or system-extracted; B management-prepared and dated; C third-party analysis; D undated, anecdotal, unattributed. Read in order of grade and relevance, not date; prior consultant decks last, because they anchor grouping. Done when every document has a grade and a read depth, and the three documents carrying most of the answer are read first. -
Extract claims (
references/templates.md§3). One assertion per row: ID (D04-C03), location (page, slide, table), close paraphrase or a short quote (under twenty words), type (fact, estimate, forecast, opinion, recommendation, gap), metric, value, period, grade, tags (sub-question, hypothesis, theme), note (definition, caveat, who benefits). Split compound sentences. Cross-reference repeats rather than re-log; repetition across independent sources is evidence, repetition of one source is not. Capture the definition with every number. Log absences as type "gap". Log both sides of a disagreement with a provisional contradiction tag; do not resolve yet. Done when every number has a period and a definition. -
Scale procedure (50+ documents). Batches of eight to twelve by triage priority; chunk long documents by section. Hold the ledger, not the corpus. Four append-only working files: index, ledger (CSV), running notes (three to six lines per batch on what changed), contradiction and gap log; layout in
references/templates.md§1–5. Consolidate every two batches: merge tags, promote recurring tags to candidate groups, prune, recount evidence per group. Controlled tag vocabulary from batch one. Delegated extraction gets the same template, tag list and a block of claim IDs; merge by ID, never by re-paraphrase. -
Group. Sort by tag and sub-question. Minto's tests per group: same kind of claim (all causes, all effects, all actions); a stated order; MECE against the sub-question. Write a summary line that states a conclusion, not a heading ("two-thirds of the decline is controllable within twelve months", not "four drivers of decline"). Reject groups whose line adds nothing beyond the claims. Done when three to six groups each have a line, a claim count and a strength rating (independent A/B sources).
-
Contradictions. Classify before resolving, in order: definitional, temporal, scope, method, genuine. Most dissolve at definition or scope. Genuine: higher grade, then more recent primary, then closer to the system. Unresolvable: log as open with claim IDs, the decision affected and the single evidence that would settle it. Never average. Each resolution gets an X-ID so the synthesis can cite it. Done when no conflict hides inside a group. Test: which contradiction, resolved the other way, would change the governing thought?
-
Pyramid. Governing thought first: one actionable sentence answering the question. Inductive structure when groups are independent lines (usual, and resilient); deductive only for a chain. Vertical logic: each line answers the reader's question about the line above. Horizontal: same kind, MECE, stated order. So-what chain on every line in three columns: finding (with claim IDs), implication for this client, action the decision-maker takes. A line stopping at finding is a summary; an action of "consider" has not been pushed. Done when the governing thought alone tells the client what to do.
-
One page and gap list. Write the SCQ last: situation contains only accepted facts, complication is the one the governing thought resolves. Governing thought, then one short paragraph per key line: line, evidence with IDs, so-what. Carry grade and type into the sentence where a line rests on an estimate or a C source. Gap list: each gap tied to the decision it affects and the evidence that closes it. Done when a reader who has seen nothing else knows what to do and why, and any challenged sentence opens its source in under a minute.
Output
Governing thought in the first 150 words. Then, each once, 1,500 to 3,500 words including tables:
- One-page synthesis (SCQ, governing thought, key lines with IDs and so-whats)
- Gap list with decisions affected
- Contradiction log (resolved and open)
- Document index
- Claim ledger (as a file or appendix; the full ledger is the audit trail, not reading matter)
No method vocabulary, checkpoint labels, framework authors, skill names or the worked example's client in the deliverable. Caveats sit in the sentence they qualify. Full example: references/worked-example.md.
Quality bar
Run silently; fix, do not report.
- Every number on the page has a claim ID with a page pointer.
- Every number in the ledger has a period and a definition.
- No group summary is a heading.
- No averaged contradiction; every open conflict is in the gap list with its consequence.
- Governing thought is actionable on its own.
- Every key line has finding, implication and action.
- Grade C and estimate-based lines carry the caveat in the sentence.
- Every gap names a decision.
- Consultant decks did not set the grouping.
- No leaked method vocabulary.
Anti-patterns
- Summary instead of synthesis: lines that describe rather than conclude.
- Silently choosing a source, or averaging two measures with the same name.
- Reading in folder order; reading the persuasive deck first.
- Compound claims in one row; numbers logged without definition.
- Holding the corpus in context instead of on disk at scale.
- Free-text tags that drift.
- A page that needs a verbal briefing to be understood.
Hand-offs
References
Frameworks: the intellectual foundation of document synthesis
3,827 wreferences/frameworks.md16 sub-sectionsTemplates and checklists for document synthesis
1,430 wreferences/templates.md14 sub-sectionsWorked example: Carrow Industrial Supply margin synthesis
4,759 wreferences/worked-example.md11 sub-sections