how to chain Claude research + ChatGPT drafting

Recommendation
Use Claude for researched extraction (facts, verbatim quotes, source snippets, metadata and a confidence score). Store that canonical research in a vector DB or cache. Hand the top-k vetted items to ChatGPT for drafting, tone refinement and multiple copy variants. Add an automated verification pass (Claude or search API) plus a human spot-check for low-confidence claims.

Decision criteria (which model for which job)
- Choose Claude when you need: long-context synthesis, careful extraction, structured JSON outputs with source snippets, or higher factuality checks. Good if output quality and traceability matter.
- Choose ChatGPT when you need: fast, polished copy, brand-tone transformations, short-turn edits, or multiple style variants. Better for iterative UX/editor workflows.
- Consider budget & rate limits: Claude extractions can be token-heavy; reduce cost by extracting only structured snippets. ChatGPT is good for many short drafts.
- Team size/workflow stage: small teams can keep manual verification; larger teams should automate verification + sampling.

Concrete orchestration and prompts (step-by-step)
1) Gather inputs: list of URLs, PDFs, raw notes. Save originals.
2) Claude — extraction (batch/one-per-source). Prompt template:
"You are a research extractor. From this source (URL/PDF/text), return JSON: {id, title, date, author, facts:[{id,text,confidence}], quotes:[{id,quote,exact_context}], sources:[{url,locator}], open_issues:[strings]}. For each fact/quote include a one-line explanation of why it's relevant and a confidence 0-100. Return only JSON."
3) Store outputs in DB (and embeddings). Deduplicate by exact quote + URL.
4) Rank & select top-K (K = 5–10 facts/quotes) by confidence + relevance score. If tokens are a concern, summarize each fact to ~50–80 chars before drafting.
5) ChatGPT — draft + tone. Prompt template:
"Input: brief audience, purpose, brand voice (3-4 adjectives), CTA, and the top-K research items (include id and source). Produce: (A) 700-word draft with inline citations like [src-id], (B) 3 headline options, (C) 2 tone variants (formal, casual), (D) short meta description. Keep quotations verbatim and cite source ids."
6) Verification pass (automated): Use Claude or a search API to verify each quoted fact: prompt Claude with the claim and ask for counter-evidence or matching independent sources and return verdict: 'verified/unverified/contradicted' with links and confidence.
7) Human spot-check: sample any 'unverified' claims for editorial review.

Rate-limit & orchestration tips (practical)
- Batch extractions per source and run parallel workers with a small concurrency limit (e.g., 4–8) plus exponential backoff on 429s. Use client-side token buckets to throttle.
- Cache everything (raw source, extracted JSON, embeddings). Avoid re-calling for unchanged URLs.
- Use progressive summarization: extract then condense to a short summary to reduce tokens sent to drafting model.
- Ask for succinct JSON outputs to simplify parsing and reduce round-trips.
- If you hit limits often, request higher quotas from the provider or spread work across time windows rather than creating additional unmanaged accounts.

Practical checklist
- [ ] Collect and archive raw sources
- [ ] Run Claude extraction job per source, store JSON
- [ ] Deduplicate + embed results in vector DB
- [ ] Select top-K facts/quotes by confidence
- [ ] Run ChatGPT drafting prompt with sources attached
- [ ] Automated verification pass (Claude/search API)
- [ ] Human spot-check low-confidence items
- [ ] Final edits and publish

Best-for / Avoid-if
- Best for: content teams that need traceable sourcing + polished brand copy. Works well when you can add a human-in-loop.
- Avoid if: you need instant single-turn output with no verification, or if you cannot tolerate the added latency/cost of two models.

If you want, I can write the exact JSON schema for Claude outputs and a ready-to-use ChatGPT prompt you can paste into your pipeline.

Compare Claude and ChatGPT