Agent Skills: do-literature-review

do-literature-review

Produce literature reviews where every citation is structurally verified before it is emitted. The skill exists to make four documented failure modes impossible to commit:

1. Phantom citations — referencing paywalled or proprietary documents that were never read.
2. Scope overreach — citing a real paper for a claim it does not actually support.
3. Author-attribution errors — wrong first author, wrong year, wrong DOI from sloppy sub-research.
4. Silent omission — listing only successful hits and hiding the searches that returned nothing.

Imperative rules ("be careful", "verify everything") do not work — agents self-report compliance falsely. Every rule below is therefore enforced by a required tool call whose result is pasted verbatim into the output, or by a required output field that cannot be filled without the tool call having happened. If you cannot fill a field, refuse to emit the citation.

Process

1. Restate and decompose — one paragraph restatement of the question + 3–8 sub-questions as a checklist.
2. Search — run queries against ≥ 2 of {arXiv, Crossref, Semantic Scholar, Google Scholar (last resort)}. Log every query string. Failed queries flow into Negative Findings.
3. Per-paper verification — see verification.md. Two mandatory tool calls per citation; raw API response pasted into the citation block.
4. Compose prose — every sourced claim points to a citation block by ID. Synthesis without a citation is marked [unsupported] inline, not deleted.
5. Negative Findings — see negative_findings.md. Every failed query and every empty sub-question listed with reason.
6. Self-audit — see self_audit.md. Re-check every citation against R1–R6 by re-reading the output, not by self-report.
7. Return in order: restatement, prose, citation blocks, Negative Findings, Self-Audit Table.

Hard rules

| Rule | Enforces | Reference | |------|----------|-----------| | R1 | Two tool calls per citation: metadata API + full-text retrieval; raw response pasted | verification.md | | R2 | Verbatim quote ≥ 25 words with page/section locator, supporting the specific claim | verification.md | | R3 | Paywall / no-full-text → metadata-only downgrade, no content claim allowed | paywall.md | | R4 | Negative Findings section is mandatory; zero-failures result is a red flag | negative_findings.md | | R5 | Self-Audit Table is mandatory; computed by re-reading the output | self_audit.md | | R6 | Authors copied byte-for-byte from the API response, not from memory or snippets | verification.md | | R7 | Refusal posture: cannot fill a required field → drop the citation, do not guess | — |

API choices and response shapes are in apis.md. Citation, Negative Findings, and Self-Audit templates are in templates.md. A worked example is in examples.md.

Anti-patterns (have happened, do not repeat)

• "Helber et al. defines land-cover classes as ..." without retrieving Helber et al. → R3 downgrade to metadata-only pointer.
• "Helber et al. 2017 discusses label reconciliation ..." when the paper has no such section → R2's verbatim-quote rule catches this.
• "Kraus et al., PLOS ONE 19(8):e0308934" when Crossref says first author is Kim → R6 forces author attribution from the API response only.
• A list of five hits with no Negative Findings → implies every search succeeded; almost never true. R4 makes this visible.

Refusal cases

If the user says "just give me a quick pointer, no need to verify" or asks for fictional/illustrative references, decline and explain that this skill is for verifiable reviews only. For single-URL summarization, route to url-reader instead.