Reference
Templates and checklists for document synthesis
Part of the Document Synthesis skill · loaded on demand from SKILL.md
Read this file when you start Stage 2 of the workflow (inventory and triage) and keep it open through Stage 8. Everything here is copy-ready. Placeholders in angle brackets are allowed in these templates and forbidden in the worked example and in any client output.
1. Question header (00_question.md)
Question: <one sentence; what the decision-maker needs to know>
Decision-maker: <role> Decision date: <date>
Sub-questions:
SQ1 <...>
SQ2 <...>
SQ3 <...>
Perimeter: <entity>, <period>, <metric definitions that anchor the work>
Hypotheses in play: H1 <...>; H2 <...>; H3 <...>
Output form: <one-page synthesis | fact base | both>
Deadline: <date> Depth: <triage | full extraction>
Assumptions made because the user was unavailable: <list, or "none">
Tag vocabulary (fixed before extraction; add only by editing this list):
<TAG> <meaning>
...
GAP a claim of type gap (something a document should contain and does not)
2. Document index (01_index.md)
| ID | Title | Author / function | Date | Type | Pages | Grade | Read depth | Batch | Status |
|---|---|---|---|---|---|---|---|---|---|
| D01 | A/B/C/D | Full / Targeted (<sections>) / Skim | 1 | Not started / In progress / Done |
Grading guide:
| Grade | Provenance | Wins a genuine contradiction against |
|---|---|---|
| A | Audited, regulator-filed, system-extracted | B, C, D |
| B | Management-prepared, internal, dated | C, D |
| C | Third-party analysis, surveys, consultant work | D |
| D | Opinion, undated, anecdotal, unattributed | Nothing; use as a pointer only |
Tie-break within a grade: more recent, then closer to the underlying system, then the source whose method is stated.
3. Claim ledger (02_ledger.csv)
CSV header row:
claim_id,doc_id,location,claim,type,metric,value,unit,period,grade,tags,hypotheses,note,logged_by,batch
Field rules:
| Field | Rule |
|---|---|
| claim_id | <doc_id>-C<nn>, numbered per document, never reused or renumbered |
| location | Page, slide, table, or section; precise enough to open in under a minute |
| claim | One assertion about one thing; close paraphrase, or a quote under twenty words in quotation marks |
| type | fact, estimate, forecast, opinion, recommendation, gap |
| metric, value, unit, period | Filled for every numeric claim; the definition goes in note |
| grade | Source grade by default; override with reason in note (e.g. "reproduces D03 p.14; A by provenance") |
| tags | From the controlled vocabulary only, semicolon-separated |
| hypotheses | H-IDs the claim bears on, with +, -, or 0 for supports, contradicts, neutral (e.g. H1-;H3+) |
| note | Definition, caveat, author's stake, cross-reference to duplicate claims, provisional contradiction flag |
Markdown rendering for reviews:
| Claim ID | Location | Claim | Type | Metric | Value | Period | Grade | Tags | Note |
|---|
Duplicate handling: when a document repeats a claim already in the ledger, do not add a row; add "also D0n p.x" to the note of the existing claim. Independent corroboration (a different primary source) does get its own row, cross-referenced both ways.
4. Running notes (03_running_notes.md)
## Batch <n>: <doc IDs> (<date>)
What changed:
- <new theme / strengthened line / contradiction / dead end>
- ...
Candidate groups and claim counts: <TAG> <n>; <TAG> <n>; ...
Open questions raised by this batch:
- ...
Governing thought, current draft: <one sentence, rewritten if it moved>
## Consolidation after batch <n>
Tags merged: <...>
Claims pruned (IDs retained): <...>
Groups promoted: <...>
Evidence count by group (independent A/B sources): <...>
5. Contradiction log (04_contradictions.md)
| X-ID | Claim A | Claim B | Class | Resolution | Reasoning | Status |
|---|---|---|---|---|---|---|
| X-01 | <claim ID: short restatement> | <claim ID: short restatement> | Definitional / Temporal / Scope / Method / Genuine | <which value is used, or "open"> | <why, with the evidence that decided it> | Resolved / Open (gap G<n>) |
Resolution order: definitional, temporal, scope, method, genuine. Never average. Every resolution gets an X-ID, and the synthesis cites the X-ID.
6. Gap list
| ID | Gap or open contradiction | Decision that waits on it | Evidence that would close it | Route |
|---|---|---|---|---|
| G1 | <what is missing> | <the decision> | <specific data or source> | Client request / boardroom-storm-research / primary-research-synthesis / market-competitive-intelligence |
Cut any gap with no decision in column three.
7. Grouping worksheet
| Group | Kind (situation / action) | Order (time / structure / degree / deductive) | Claims | Summary line, first draft | Summary line after blank-assertion test | Independent A/B sources |
|---|
Tests to run on each row before it becomes a key line:
- Same kind: all members are the same type of idea.
- Order named: the order column is filled and would be recognisable to a reader.
- MECE against the parent: no claim fits two groups; no obvious category is missing.
- Summary not blank: the line states a conclusion, not a category.
- Summary adds information: the line says something no single claim says.
8. So-what chain
| Key line | Finding (claim IDs) | Implication for this client | Action (verb, object, owner or threshold, time) |
|---|
Ask "so what?" once more of each implication. If a new action appears, replace the first.
9. Pyramid (05_pyramid.md)
Question: <from 00_question.md>
Horizontal logic: <inductive | deductive>
Governing thought: <one sentence; answers the question; contains a number and an action>
Key line 1: <summary line>
Evidence: <claim IDs, X-IDs>
So what: <action>
Key line 2: ...
Key line 3: ...
(Key lines 4 and 5 only if each adds a distinct kind of support)
If deductive:
Premise 1: <...> [IDs]
Premise 2: <...> [IDs]
Therefore: <...>
10. One-page synthesis (06_synthesis.md)
# <Title that states the answer, not the topic>
Situation: <one or two sentences the reader already agrees with>
Complication: <what changed or conflicts, making the question urgent>
Question: <the question>
Answer: <the governing thought>
1. <Key line 1>. <Evidence with claim IDs; grade caveats where a line rests on C or D.> So what: <action>.
2. <Key line 2>. ...
3. <Key line 3>. ...
(4., 5. if needed)
<Bridge or reconciliation table if the question is quantitative>
Contradiction and gap list: G1 <...>; G2 <...>
Attachments: document index (<n> documents), claim ledger (<n> claims), contradiction log (<n> entries)
Length: the body fits on one page; attachments are separate.
11. Batch brief for delegated extraction
Use this when handing a block of documents to a sub-agent or analyst.
Question and sub-questions: <paste from 00_question.md>
Your documents: <doc IDs and files>
Claim ID block: <doc IDs are yours; number claims per document from C01>
Tag vocabulary: <paste; do not add tags; note candidates in your batch note>
Ledger template: <paste CSV header>
Rules:
- One claim per row; definitions with every number; type recorded honestly.
- Quote under twenty words for anything contentious.
- Log gaps (type gap) where a document should contain something and does not.
- Do not resolve contradictions; flag them in the note field with "CONTRA <other claim ID>".
Return: ledger rows, a batch note in the running-notes format, and candidate tags.
12. Pre-delivery checklist
Run in order; stop at the first failure and fix it.
- The governing thought answers the question in 00_question.md as written.
- The governing thought contains a number and an action, and names no category.
- Every key line passes the blank-assertion test.
- Under each parent, children are the same kind and in a named order.
- For inductive pyramids, removing any key line leaves the governing thought standing.
- Every number in the synthesis has a claim ID; five random numbers were traced to a page.
- Every key line that rests on grade C or D says so in the text.
- Every contradiction in the log is resolved or appears in the gap list; none was averaged.
- Every gap names the decision that waits on it.
- At least one key line states something no single document says.
- Claim IDs in the synthesis exist in the ledger (grep each ID).
- No placeholders, no em dashes, no emojis in client-facing text.
- Hand-off package prepared for
executive-communication: pyramid, synthesis, ledger, index.