Recommendation (short):
Use Claude as the “analysis engine” to extract structured findings, evidence, and data tables; hand off machine-readable structured output (JSON) to ChatGPT for final editorialization, tone, and document-level polishing. Keep a human QA step before final export.

Orchestration steps (practical):
1) Ingest & chunk: collect sources (PDFs, papers, datasets). Convert to text and chunk into 2–5k token windows. Store chunks in a vector DB (optional) or S3.
2) Analysis pass (Claude): for each chunk, ask Claude to extract claims, evidence, data points, and citations. Request a strict JSON schema: e.g. {section_id, claim[], evidence[], confidence_score, raw_text_ref}.
3) Synthesis pass (Claude): ask Claude to synthesize the chunk-level JSONs into a structured outline + detailed findings + recommended figures. Output as JSON with keys: executive_summary, outline, sections[], appendix[].
4) Editorial pass (ChatGPT): feed the Claude JSON into ChatGPT with a prompt to produce the final report in your target tone (10k-word target), keeping sections and citations intact. Ask for Markdown and a DOCX-ready structure.
5) QA & human edit: run automated checks (citation links, duplicate facts), then have an editor review. Optionally run a final pass with ChatGPT for copyediting.
6) Export: render to DOCX/PDF/HTML and archive the Claude JSON and intermediate artifacts.

Prompt handoffs (concise examples):
- Claude (analysis chunk prompt): “Extract claims and evidence from this chunk. Return strict JSON with fields: id, claims[ {text, span_start, span_end}], evidence[{quote, source, page}], confidence(0-1). No extra commentary.”
- Claude (synthesis prompt): “Given the array of chunk JSONs, synthesize a report-outline JSON: executive_summary, sections[{id,title,bullets,figures}], appendix. Keep citations linked to source ids.”
- ChatGPT (editorial prompt): “Using this report JSON, write a 10,000-word report in [tone], preserving section structure and citations. Produce Markdown and a DOCX-compatible structure (H1/H2/H3). Add transitions, human examples, and a short abstract. Include a references section that maps source ids to full bibliographic entries.”

File formats & handoff conventions:
- Chunk output: JSONL (one JSON per chunk) + original text refs.
- Synthesis & analysis: canonical JSON with fixed schema (export as .json). Use explicit keys for sections, claims, figures.
- Editorial output: Markdown (.md) for easy editing + DOCX (.docx) for clients + PDF for distribution.

Zapier/Make glue & example orchestration:
- Zapier-style flow: Webhook -> Store chunks in Google Drive (or S3) -> HTTP request to Claude API (analysis) -> Save JSON to Airtable/Drive -> HTTP request to ChatGPT/OpenAI API (editorial) -> Save Markdown/DOCX -> Slack/email notification.
- Small Node.js pseudo-code (outline):
1. fetch(claudeEndpoint, {prompt: analysisPrompt}) -> parsedJSON
2. save(parsedJSON)
3. fetch(openaiEndpoint, {prompt: editorialPromptWithJSON}) -> finalMarkdown
4. convert Markdown to DOCX (pandoc or mammoth) -> upload

Decision criteria (when to do hybrid):
- Best-for: deep research that needs careful citation, long context, and structured evidence extraction. Works well with multi-person teams (analysts + editors).
- Avoid-if: short blog posts, tight budgets, or when you need realtime low-latency answers (single-model is faster/cheaper).
- Depends on: budget (two APIs cost more), team size (smaller teams may skip multi-pass), and stage (early exploration vs final delivery).

Practical checklist before first run:
- Define JSON schema and examples.
- Build chunking & OCR pipeline; validate chunk size.
- Create Claude analysis prompt templates and test on 5 documents.
- Create ChatGPT editorial template (tone guide, length targets).
- Implement storage (S3/Drive), versioning, and human QA step.
- Instrument costs & latency; run a pilot on 1 report.

Quality note: this hybrid approach trades cost and latency for stronger factual structure + polished human tone. If you want, I can draft the exact JSON schema and the three concrete prompt templates to paste into Claude and ChatGPT.

Compare ChatGPT and Gemini