Problem summary
Claude can produce convincing but untraceable citations when you ask it to synthesize dozens of papers in one go. The reliable fix is not a single magic prompt — it’s a reproducible pipeline and a few strict prompt rules that force provenance to be explicit and machine-checkable.

Recommendation (short)
Use a 3‑pass pipeline: Extraction → Synthesis with enforced inline source tags → Automated Verification. Store and pass only vetted snippets (with doc IDs, page/offsets, URLs/DOIs) and require Claude to refuse claims it cannot link to those snippets.

Why this works
Claude is strong with long context, but hallucinations happen when it invents links while free‑generating. By converting each paper into compact, verifiable units (quotes + metadata) and by making citations mandatory tokens (not free text), you remove the model’s freedom to invent sources.

Concrete prompt patterns
- Extraction prompt (for each doc):
"Return JSON for DOC_ID with fields: metadata (title, authors, year, DOI/URL), 3 verifiable quotes with exact page/paragraph and character offsets, a 1‑sentence summary, and 5 keywords. If you cannot find a quote, return an empty list."
- Synthesis prompt (section-level):
"Write Section X using ONLY the provided snippets. Every factual claim must end with a citation token of the form [DOC42:p3:char100-175]. If you cannot support a claim from a snippet, write [UNVERIFIED] and do not invent a source."
- Verification prompt (post‑synthesis):
"For each sentence in Section X, return the supporting snippet(s) (DOC_ID + exact quote + location) that justify it, or mark MISSING."

Doc‑chaining / context handling
- Precompute a provenance table (CSV/JSON) with one row per snippet: DOC_ID, author, year, page, offset, DOI/URL, snippet text. Pass this table (or cyclically the most-relevant rows) into Claude rather than full PDFs.
- Chunk long docs into overlapping passages and assign stable IDs (DOC1_P001, DOC1_P002) so citations reference passages, not vague pages.
- For very long reports, synthesize per-section then stitch: keep the provenance table as the single source of truth you feed into each section job.

Checklist (practical)
1. Ingest: convert PDFs → snippets with page/offset + DOI/URL.
2. Extract: run Claude to produce 3+ verifiable quotes per doc (store as JSON).
3. Index: build a small vector/index so you can retrieve top-N snippets per claim.
4. Synthesize: ask Claude to write using only provided snippet IDs and strict citation tokens.
5. Verify: run automated pass to match each sentence to snippet(s); mark MISSING/UNVERIFIED.
6. Human QC: spot-check randomly and for critical claims, open source URL or PDF to confirm.
7. Fix: For MISSING items, either remove claim, downgrade language, or add correct snippet + resynthesize.

Decision criteria
- Use strict pipeline when you need high reliability (papers, regulatory, policy). Requires more engineering and time.
- Looser approach (free paraphrase + human review) is OK for early drafts or brainstorming.
- Teams with low budget can start manually: extract 1–2 quotes per doc. Larger teams should automate extraction + indexing.

Best‑for / Avoid‑if
- Best for: literature reviews, academic reports, compliance documents, grant writing.
- Avoid if: you only need fast sketches, have no time for verification, or the output can tolerate citation noise.

Tool note
Claude’s long‑context capability makes it a good fit for the multi‑pass, provenance table approach. Use Claude for extraction+synthesis; keep a final human or automated checker for verification.

Quick final tip
Make the model’s failure mode explicit: require the token [UNVERIFIED] for any unsupported claim and refuse to continue if >X% of sentences are UNVERIFIED. That simple rule drastically reduces invented citations.

Compare Claude and ChatGPT