Recommendation
Use a chunk-then-synthesize workflow: break the 30k-word brief into topic chunks you process one-by-one (with overlap and strict source capture), then run staged synthesis passes that assemble, reconcile, and audit citations into the final deliverable. For best long-context reliability prefer Claude (longer context / careful analysis); ChatGPT works as the default model for smaller budgets or iterative human-led editing.

Decision criteria (which approach to pick)
- Budget: paid Claude/ChatGPT+ for larger context and speed; otherwise use smaller chunks and more synthesis cycles.
- Skill level: single writer can run chunk→synth; a multi-person team can parallelize chunking and fact-checking.
- Workflow stage: heavy primary research requires more verification passes and human interviews; late-stage polishing needs style and coherence passes.
- Desired output quality: high-stakes reports require human audit of all primary claims and links.

Prompt-chunking strategy (practical)
1) Create a master outline (30–50 headings; executive summary, market sizing, segmentation, competitive landscape, channels, pricing, risks, recommendations).
2) Define chunk size: aim for 1,200–2,500 words per chunk (or 1,000–2,000 tokens) so verification is tractable. For 30k words that’s ~12–25 chunks.
3) Number and name chunks: CH01_Market_Definition, CH02_TAM_SAM, etc. Always include chunk number in prompts.
4) Overlap: include 200–400 words overlap between adjacent chunks to preserve thread continuity and prevent context gaps.
5) Chunk prompt template: provide the chunk name, precise instructions, required outputs (summary, bullet facts with citations, outstanding questions). Example: “Process CH05_Competitive_Landscape. Output: (A) 600–800word draft for this section; (B) a bullet list of 12 provenance-tagged facts [sourceID, title, author, URL, date]; (C) 3 remaining evidence gaps.”

Citation policy (practical, auditable)
- Required metadata per source: sourceID, title, author, publisher, URL, date, type (report, article, dataset, interview).
- In-chunk citations: attach sourceIDs inline after facts (e.g., [S12]). Do not synthesize unattributed facts.
- Primary vs secondary: flag primary sources (datasets, interviews, regulatory filings). Prioritize primary for quantitative claims.
- Verbatim quotes: include exact quote and sourceID; mark with quotes.
- Verification step: after synthesis, run an automated link-check and a human spot-check on 10–20% of claims.
- Output bibliography: numbered reference list in chosen style (APA/Chicago) appended to the report and each chapter.

Synthesis steps (staged)
1) Merge pass: combine chunk drafts in outline order, resolve overlapping material, keep provenance lists concatenated.
2) Reconciliation pass: run prompts to detect contradictions across chunks (list contradictions and preferred source).
3) Fill-gaps pass: issue targeted research prompts for remaining evidence gaps.
4) Coherence & tone pass: normalize voice, remove repetition, craft executive summary and recommendations.
5) Final audit: verify critical claims, check links, produce methodology/limitations appendix and a provenance ledger.

Checklist (practical)
- [ ] Create master outline and chunk list
- [ ] Assign chunk prompts and naming convention
- [ ] Capture full source metadata for every cited item
- [ ] Maintain overlap and chunk numbering
- [ ] Run merge, reconciliation, fill-gaps, and tone passes
- [ ] Perform automated link checks + human spot-checks
- [ ] Produce bibliography, methodology, and provenance ledger

Best-for / Avoid-if
- Best for: teams needing traceable, long reports with verifiable claims. Claude is best for long-context heavy synthesis; ChatGPT is fine for iterative drafting and editing.
- Avoid if: you have no capacity for human verification or very tight budgets—don’t skip the verification step.

If you want, I can draft the master outline and ready-to-run chunk prompts for Claude (recommended) or ChatGPT next.

Compare Claude and ChatGPT