Given N PDFs on the same subject:

1. Inventory every file (pages, title, structure, extraction risks).
2. Extract every substantive idea from every page of every file, tagging each I001, I002, … with source (D1, D2…) and page.
3. Build canonical idea records, merging duplicates while preserving distinct source links.
4. Compare across documents: consensus, unique contributions, duplication, contradictions, gaps.
5. Assign a disposition to every idea: INCLUDE | MERGE | ADAPT | VALIDATE | DEFER | REJECT | UNRESOLVED.
6. Write the master document organized by topic, not by source document.
7. Build comparison matrix, idea register, conflict log, and provenance index.
8. Run coverage, consistency, and PDF-rendering audits.
9. Deliver: Master PDF, Comparison Matrix, Idea/Decision Register, Completion Report.

Never summarize documents one at a time and concatenate — that is the failure mode this skill exists to avoid.

Progress:
- [ ] Stage 1: Input inventory (files, pages, structure, integrity issues)
- [ ] Stage 2: Full extraction (every page, every document, no exec-summary shortcuts)
- [ ] Stage 3: Idea normalization (canonical records, taxonomy adapted to subject)
- [ ] Stage 4: Cross-document comparison (consensus/unique/duplicate/contradiction/gaps)
- [ ] Stage 5: Critical evaluation & disposition (every idea gets one status)
- [ ] Stage 6: Master synthesis draft (organized by topic/decision, not by source)
- [ ] Stage 7: Build full PDF structure (exec synthesis + analysis + audit appendices)
- [ ] Stage 8: Traceability pass (every claim cited [Source: Dx, p.y])
- [ ] Stage 9: QA gates (coverage, consistency, PDF production)
- [ ] Stage 10: Deliver outputs + completion report

Stage 1 — Inventory

Record per file: filename, source ID (D1…Dn), page count, title, apparent purpose, structural elements (tables/figures/appendices), and any extraction limitation (scanned pages, images, unreadable sections). Do not proceed to synthesis until this is complete and gaps are flagged.

Stage 2 — Extraction

Read every accessible page. Capture arguments, recommendations, evidence, assumptions, examples, trade-offs, risks, terminology, tables/diagrams, footnoted caveats. Assign idea IDs with source + page. Paraphrase accurately; quote only when exact wording matters. Treat footnotes and table cells as first-class content, not noise. A conditional recommendation and its unconditional cousin are different ideas — don't flatten them.

Stage 3 — Normalization

Build a taxonomy fit to the subject (e.g., strategic principles, recommendations, risks, metrics, assumptions, open questions — adapt freely). Create one canonical record per distinct concept with links to all original occurrences. Same label ≠ same idea; different label ≠ different idea. Verify meaning, not wording.

Stage 4 — Cross-Document Comparison

Run five explicit passes:

• Consensus: which docs share an idea, do they agree on substance or just language, does agreement imply correctness (usually no)?
• Unique contributions: ideas in only 1–2 docs — never discard for rarity alone.
• Duplication: exact / semantic / complementary / similar-but-different / superficially-related-but-different-problem.
• Contradiction: opposing claims, who holds which position, why, and whether resolvable from evidence given.
• Missing context: shared blind spots and gaps one doc raises that others ignore.

Stage 5 — Evaluation & Disposition

Judge each idea on strategic alignment, expected value, evidence quality, feasibility, risk, novelty, complementarity — qualitatively, no invented numeric precision. Assign exactly one disposition per idea. Rejected/deferred/unresolved ideas stay visible in the audit trail — never delete silently. When ideas compete for limited resources, make the trade-off explicit instead of recommending everything.

Stage 6 — Master Synthesis

Organize by topic/problem/decision. Lead with strongest conclusions. Merge overlapping recommendations into one coherent explanation while keeping each source's distinct reasoning where it adds value. Fold in unique insights that earn their place. Flag trade-offs and disagreements openly. Keep dependencies and caveats attached to their claims. Label content as source-derived, synthesized, or inferred.

Stage 7 — Structure the PDF

Part I — Executive synthesis: cover page; executive overview; key findings across all docs; recommended master strategy. Part II — Detailed analysis: topic chapters with full depth; critical comparisons/trade-offs; risks/assumptions/evidence gaps; implementation/action plan (mark estimates as proposed, never invent budgets/dates). Part III — Audit appendices: master idea register; source-by-source contribution audit; cross-document comparison matrix (split by topic if large); conflict/decision log; provenance index (idea ID → file/page/section).

Stage 8 — Traceability

Cite as [Source: D3, p.12]; list all sources for multi-origin ideas. Tag every claim: source-derived / synthesized / inferred / externally verified (only if research was explicitly authorized). Never let an inference masquerade as something a source stated.

Stage 9 — QA Gates

Coverage: every file inventoried, every page examined, every idea has a canonical record + disposition, every unique contribution addressed, every contradiction logged, nothing dropped for appearing in only one doc. Consistency: no redundant repetition across chapters, no false-merge from similar wording, no buried contradictions, caveats still attached, executive summary matches the detailed register, quantitative claims faithfully represented. PDF production: renders correctly, proper heading hierarchy, TOC and page numbers, tables fit page bounds, nothing clipped/duplicated/missing, clean page breaks for long tables, searchable text where possible.

If any gate fails, fix and re-run that gate before delivery.

Stage 10 — Deliverables

• A. Master Synthesis PDF
• B. Cross-Document Comparison Matrix (spreadsheet/table preferred)
• C. Master Idea & Decision Register
• D. Completion & Coverage Report (files/pages processed, limitations, unresolved issues, QA results)

Only claim a file exists if it was actually produced and verified.

Example 1: Input: 6 PDFs from different LLMs proposing a company's AI adoption strategy. Four recommend a centralized AI platform team; two recommend fully decentralized ownership; one uniquely proposes a federated hybrid model with a central guild plus embedded practitioners. Output: Master doc presents the federated hybrid as the lead recommendation, synthesizing why pure centralization (4 docs) and pure decentralization (2 docs) each fail on a specific axis (speed vs. governance), with the unique idea elevated to INCLUDE status, full rationale cited [Source: D5, p.7], and the centralization/decentralization debate logged in the conflict log as resolved-by-synthesis rather than majority vote.

Example 2: Input: 7 PDFs analyzing a market-entry decision; three cite a "40% CAGR" figure from the same underlying claim reworded differently, one cites 25% CAGR from a different line of reasoning. Output: Treated as a factual contradiction, not a wording difference. Both figures preserved with their source and reasoning, disposition UNRESOLVED pending external verification, explicitly flagged in Part II risk section and the conflict log — not silently averaged to "~35%."