Recommendation (short):
Use Claude for long-context extraction and a two-stage prompt flow: 1) chunk+annotate the 40‑page report into coherent sections and ask Claude to build an evidence map, 2) synthesize the evidence map into slide‑ready bullets (3–5 bullets/slide, 6–8 words per bullet). That preserves traceability (page refs) and yields slide-ready output.

Why this works:
Claude handles long-context documents well, so you can feed larger chunks and ask for careful, sourced extraction. The two-stage flow keeps outputs auditable and lets you iterate on tone/length for slides.

Decision criteria (which template to use):
- Use Full‑Report Template when report length ≤ 40 pages and you can feed entire text (best for accuracy).
- Use Chunk+Merge Template when reports are scanned or >40 pages or you want parallel processing (best for throughput and team pipelines).
- Choose single-shot slide template only when you trust the LM’s summarization and need quick results (lower auditability).
Factors: budget (API costs scale with input size), skill level (chunking requires light dev work), team size (pipelines benefit large teams), output quality needs (high quality → more iterations and human validation).

Practical checklist before prompting:
- OCR clean (if scanned) and remove extraneous pages.
- Split into logical sections (exec summary, market sizing, product features, financials, go‑to‑market, appendix) or fixed chunks ~3,000–5,000 words.
- Number pages and retain original page numbers in each chunk.
- Decide output format: JSON for downstream tools or slide bullets for analysts.

Tested prompt templates (plug-and-play)
1) Chunk-level extraction (run for each chunk):
System: "You are an analyst. Extract competitor mentions, claims, metrics, strengths, weaknesses, strategy, and supporting evidence with page numbers. Return JSON: {competitor, insight_type, summary, evidence_pages, verbatim_quote_if_any}. Keep each value short."
User: "CHUNK_X/Y (pages A–B):n{{chunk_text}}nReturn only valid JSON array of findings."

2) Merge + evidence map (after collecting arrays):
System: "You are a senior market analyst. Merge the arrays into a deduplicated evidence map. Group by competitor and insight_type, combine evidence_pages, and flag confidence (High/Med/Low) with rationale. Output JSON."
User: "Here are findings from chunks: {{all_findings}}. Merge and dedupe."

3) Slide‑ready bullets (from merged map):
System: "You are an executive presentation writer. Create slide titles and 3–5 bullets per slide. Bullets must be short (6–12 words), action‑oriented, and include a bracketed citation like [p.12]. Provide recommended visual (chart/table) per slide."
User: "Produce slides for top 6 competitor insights: {{merged_map}}. Output as markdown-like slide blocks."

Best-for / Avoid-if
- Best for: analysts needing traceable competitor insights from long reports, and teams that will validate outputs.
- Avoid if: you need legally binding citations (use human verification) or you cannot preprocess/number pages.

Quick validation steps after generation:
- Spot‑check 5 random bullets back to the original pages.
- Flag Low‑confidence items for human review.
- Iterate tone/length with Claude by asking for ‘shorter’ or ‘more assertive’ variants.

If you want, I can provide ready-to-run API-friendly prompt JSON (chunking loop + merge + slide templates) tailored to your pipeline and desired slide count. CTA tool: claude

Compare Claude and ChatGPT