How to chain ChatGPT with Claude for summarization

Recommendation (short):
Use Claude to ingest and chunk long documents, produce chunk-level extractive notes with provenance, then hand a compact, provenance-rich JSON to ChatGPT to produce a short, punchy executive summary. Keep provenance attached as a small JSON manifest (chunk IDs, source offsets/URLs, ingestion timestamp, model & prompt used, confidence tag) so the summary can link back to originals.

Why this works
- Claude handles long context and careful analysis well (good for chunking, extraction, citations).
- ChatGPT is excellent for concise, readable executive prose.
- Passing structured provenance prevents losing source traceability when you rephrase or condense.

Decision criteria (pick strategy by need):
- Use full Claude-first pipeline when: documents >50K tokens, need high-fidelity extraction, multiple source types, or compliance/audit requirements.
- Skip deep chunking if docs are short (<5K tokens) or speed/cost is critical.
- If you need verbatim quotes included, keep extractive snippets in the provenance manifest.
- Team size & skill: if you have engineers, implement JSON/Vector store; for single users, copy/paste structured JSON between tools is OK.
- Budget: two-model pipeline costs more; prefer single-model if budget is tight.

Practical checklist (step-by-step):
1) Ingest with Claude
- System prompt: ask Claude to chunk the document into ~2–4 KB logical chunks (preserve paragraph boundaries) and for each chunk emit: chunk_id, text, start_offset, end_offset, source_url/page, ingestion_time, confidence_est.
- Also request a short extractive summary (1–2 sentences) and 3–5 salient quotes (with offsets) per chunk.
- Output format: JSONL or a single JSON array.
2) Store chunks + embeddings (optional)
- Save chunk JSON to your datastore; compute embeddings if you want retrieval (store vector_id per chunk).
3) Build a provenance manifest (JSON) including: doc_id, original_url, ingestion_model:Claude+version, ingestion_prompt_hash, chunk_ids[], chunk_summaries[], timestamp.
- Example schema (keys): doc_id, source, chunks:[{id, start, end, summary, quotes:[{text,offset}]}], ingestion_meta:{model,prompt,ts}.
4) Hand a trimmed payload to ChatGPT
- System prompt example: “You are an executive summarizer. Given the provided provenance manifest and the chunk summaries below, produce a 3-paragraph executive summary (max 180 words). At the end, list the top 3 source chunk_ids and a 1-line rationale for each citation.”
- Payload to ChatGPT: include the provenance manifest and only the chunk summaries + a couple of full quotes (not whole chunks) to keep token usage low.
5) Attach final provenance
- Append ChatGPT’s output to the manifest as summary_text, summarizer_model:ChatGPT+version, prompt_hash, created_at.

Provenance JSON example (condensed):
{doc_id, source_url, ingestion:{model,ts,prompt_hash}, chunks:[{id,summary,quote_offsets}], summary:{model,ts,text, cited_chunk_ids:[...]}}

Best-for / Avoid-if
- Best for: multi-page reports, research reviews, and compliance-required audits where traceability matters.
- Avoid if: you need ultra-low latency or have strict single-model cost constraints.

Final tips
- Keep prompts versioned (hash or small identifier) and save them with the manifest.
- If you want automatic retrieval later, store chunk embeddings and chunk_id keys.
- If you prefer GUI tools, export/import the JSON between Claude and ChatGPT via clipboard or storage bucket.

If you want, I can give exact system + user prompt templates and a ready-to-copy JSON schema for your pipeline.

Compare ChatGPT and Gemini