literature-review

Literature Review

Produce a cited literature review grounded in primary sources — papers, preprints, lab pages, author profiles — not secondary write-ups. Derive a short slug (lowercase, hyphens, ≤5 words) from the topic and use it for every file in the run.

The heavy searching is not this skill's job — hand it to a background agent. Run /deep-research or /research to sweep sources, then do the academic synthesis layer here. Keep intermediate notes in the scratchpad; write the final artifact where the project keeps research notes (match the existing convention; if none, pick a sensible research/-style location and say where).

Workflow

1. Plan — Scope the review: key questions, source types (papers, web, repos), time window, expected sections. Summarise the plan to the user in one breath and continue — do not wait for confirmation unless they asked to review it first. When the input names a lab, PI, author, or profile, run it as a publication-corpus review: resolve the identity first, collect the reachable publication list (titles, years, venues, DOIs/URLs, gaps), then map the trajectory across that corpus.
2. Gather — Spawn the researcher subagent for wide sweeps (or /deep-research); search directly for narrow ones. Check the user's own library first with /zotero-search (what's already collected, with local PDFs and citekeys), then cross-source the web with WebSearch and /scholar-search (search … --source all, resolve, cited-by, refs) so metadata comes with real DOIs and citation counts. Prefer stable source URLs (arXiv, OpenAlex, journal DOIs, lab pages). Mark every planned question done, blocked, or superseded — never silently drop one.
3. Synthesize — Separate consensus, disagreements, and open questions. For a corpus review, name 3–5 research trajectories and the 3–5 papers that most changed the corpus direction, ranked by contrastive originality and methodology strength rather than author prestige. Use a Mermaid taxonomy / trajectory map or a comparison table only when the structure is source-supported. Propose concrete next reading or experiments.
4. Cite & verify — Spawn the verifier subagent to add inline citations and check every source URL resolves, then the reviewer subagent for an adversarial pass on unsupported claims, logical gaps, and single-source critical findings. Fix fatal issues before delivering; record major ones under Open Questions.
5. Deliver — Save <slug>.md and a <slug>.provenance.md sidecar listing: date, sources consulted vs. accepted vs. rejected, verification status, and the intermediate files used (plus the publication-log path and unresolved gaps for a corpus review). Confirm both files exist on disk before you stop.

Adapted from companion-inc/feynman (MIT). See ../ATTRIBUTION.md.