Literature synthesis

Prompt blocks for infrastructure.llm or project synthesis.py after literature search.

Natural invoke

• "Synthesize these 12 papers into thematic clusters with citation keys"
• "Per-paper structured notes: contribution, method, evidence, limitation"
• "Gap analysis: what experiments would close open questions?"

Inputs to confirm

• Corpus — list of papers (title, authors, year, abstract, optional fulltext).
• Goal paragraph — for gap-analysis variant only.

Paper block format

Each paper renders as:

text
### {citation_key} — {title} ({year})
**Authors:** {authors}
**DOI / URL:** {doi or url}
**Abstract:** {abstract}

Workflows

Per-paper analysis

Structured note: CONTRIBUTION (one sentence), METHOD (2–3 bullets), EVIDENCE (2–3 bullets), LIMITATION (one bullet), TAGS (3–7 lowercase). Cite as [{citation_key}].

Cross-paper synthesis

Group into 3–7 thematic clusters; summarise dominant approach; agreements/disagreements; 3 open questions. Cite only corpus keys in square brackets.

Gap analysis

Given GOAL + corpus: per paper COVERS/LACKS sub-claims; propose 3 follow-up experiments. Cite by key.

Style

• Bracket-key citations only — not bare titles.
• Bound input: truncate fulltext; temperature=0.0, seed=42 for replay.
• Optional second pass: infrastructure.llm.validation.validate_complete.

Verification

• Outputs under output/llm/ or project convention; keys match references.bib.
• See projects/templates/template_search_project/ for wired implementation.

When NOT to use

• Running literature search backends → infrastructure/search/ and project search scripts
• New manuscript from scratch → manuscript-creation

References

Full prompt text and programmatic example: references/prompt-blocks.md