Recommendation (one-line)
Use Claude to synthesize long context into a concise claims+evidence bundle, then run an automated ChatGPT cross-check that verifies each claim and URL before publishing.

Why this works
Claude handles long context and careful analysis well; ChatGPT is fast and reliable for targeted verification prompts. This keeps the workflow light, API-driven, and auditable without full manual fact-checking for low- to medium-risk publications.

Decision criteria — when to use this vs heavier options
- Use this workflow when: tight deadlines, moderate legal/risk exposure, small team, outputs are summaries or research briefs.
- Avoid if: claims could trigger legal action, you need formal fact-check sign-off, or you produce journalistic/investigative pieces—then add human fact-checkers or paid verification services.
- Budget/skill tradeoffs: lightweight automation is cheap but needs prompt engineering and a dev to wire Claude→ChatGPT. Larger teams may prefer a dedicated QC person.

Practical checklist (publish only when all boxes checked)
- [ ] Every factual claim has a supporting URL and source snippet saved.
- [ ] All URLs return 200 and match the cited claim (or flagged otherwise).
- [ ] ChatGPT verdict for each claim = “Verified” or has a recommended correction.
- [ ] Any flagged items resolved or explicitly marked in the published piece.
- [ ] Final summary includes provenance footnotes (source id + timestamp).

Lightweight workflow (step-by-step)
1) Ingest & synthesize (Claude):
- Feed full docs or scraped content into Claude in chunks. Ask for: concise summary, list of discrete claims (numbered), the exact sentence in-source that supports each claim, and the source URL + snippet. Output a JSON with fields: claim_id, claim_text, source_url, source_snippet.
2) Bundle for cross-check (transform):
- Convert Claude JSON into a checklist for ChatGPT: each claim with its snippet and URL.
3) Cross-check (ChatGPT):
- Prompt ChatGPT to verify each claim against the provided snippet/URL and to: return status (Verified/Unverified/Needs Context), confidence (0–100%), suggested correction, and URL health (OK/Redirect/404/Paywall). Ask ChatGPT to provide 1–2 corroborating sources if available.
4) Apply fixes & publish:
- Automatically apply ChatGPT-suggested textual edits or flag for human review. Include provenance footnotes in the final summary.

Minimal prompt templates (start here)
- Claude (synthesis): "Extract discrete factual claims from the text. For each claim output: id, concise claim_text, exact source sentence, source_url, and a 1-line summary of why it supports the claim. Return JSON array."
- ChatGPT (cross-check): "For each claim and its source_snippet+URL, say: status (Verified/Unverified/Needs Context), confidence 0–100, reason, suggested correction (if any), and url_health. If unverified, list up to 2 alternative sources."

Best-for / Avoid-if
- Best for: research briefs, newsletters, internal summaries, low-risk public content.
- Avoid if: high-stakes journalism, legal/regulatory claims, or when you need human-certified verification.

Implementation tips
- Store Claude output as structured JSON so ChatGPT can iterate reliably. Use retries for transient URL errors. Log decisions for audit trails.

If you want, I can draft exact API call payloads and a small prompt-testing checklist for your current sources/tooling.

Compare Claude and ChatGPT