agentleFS
arthur-plautzarthur-plautz/research-toolkit/.claude/skills/literature-review/SKILL.md

literature-review

Skill0 starsChanged 4 months agoMITView on GitHub

Runs a structured literature review on a given topic — refines scope through interrogation, searches broadly for ~10 strictly relevant sources with direct application to the discussion, persists them via /fetch-paper and /references, and writes a synthesis document under docs/reviews/. Also used to view or edit existing reviews.

92 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use pdf-explore, paper-narrative and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • JimLiuSkill

    pdf-explore

    Use this skill when the user has attached a PDF, paper, report, or other document and the answer needs content from more than one place in it: summarize the methods or any other section, compare sections, find where a topic is discussed, read a value or label off a figure or chart, or find/list/extract every instance of something across the whole document (datasets, benchmarks, citations, figures, table rows, accession numbers — including appendices). Skip it only for a single lookup of 1–4 pages quoted in your very next response — `read_file(pages=[...])` attaches pages as images that are dropped from context after one turn, so multi-section answers end up re-reading the same ranges repeatedly. Parses the PDF once in the Python kernel: `pdf_pages` (pages as persistent text), `pdf_outline` (TOC), `pdf_scan` (rank pages by relevance), `pdf_map`/`pdf_extract` (per-page summary / structured fields via parallel haiku calls). For PDF creation/manipulation, use reportlab/pypdf directly.

    Also in 9 projects that use this

  • JimLiuSkill

    paper-narrative

    Judge and reshape the STORY a paper's figures tell. Input is the work itself — manuscript (or abstract) + figure deck — no hand-written brief. `derive_paper_brief(abstract, captions)` extracts pitch/vision/per-figure-claims; a handling-editor reviewer on the full deck returns hook_verdict (would Fig 1 make me send this for review?), arc (hook→mechanism→evidence→application), figure_moves (panels in the wrong figure), missing_panels (concrete analyses to RUN), kill_list, and boldest_defensible_fig1. Hands per-figure claims to `figure-composer`. Load when writing or revising a paper.

    Also in 5 projects that use this

  • Norman-burySkill

    brainstorming-research

    Use before writing any paper - clarifies paper type, discipline, topic, methodology, and chapter structure through natural dialogue

    Also in 4 projects that use this

  • lingzhi227Skill

    latex-formatting

    Handle LaTeX formatting, templates, and styling for academic papers. Set up conference templates (ICML, ICLR, NeurIPS, AAAI, ACL), fix formatting issues, manage packages, and ensure venue-specific compliance. Use when the user needs to set up a paper template, fix LaTeX formatting, or prepare for submission.

    Also in 4 projects that use this

  • lingzhi227Skill

    novelty-assessment

    Assess research idea novelty through systematic literature search. Multi-round search-evaluate loops with harsh critic persona. Binary novel/not-novel decision with justification. Use before committing to a research direction.

    Also in 4 projects that use this

  • +6 more from JimLiu
  • +1 more from lingzhi227

The file

We split this skill into the rules it gives an agent

11 rules. 0 are common, 2 are distinctive, and 9 are very specific to this project.

  • 0

    Common

    Most of the 72 edited versions we read keep these.

  • 2

    Distinctive

    This author's own choices. Worth a look.

  • 9

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Ask targeted questions — do not ask for anything inferable from context (e.g.

    line 39
  2. Distinctive:

    Ask: "Proceed with this search?"

    line 62
  3. Very specific:

    Search broadly (WebSearch, Semantic Scholar/OpenAlex queries as used by /fetch-paper's resolution phase) using the confirmed terms.

    line 70
  4. Very specific:

    Check docs/references/INDEX.md for candidates already covered — keep them in the shortlist (they still count toward the 10) but they will …

    line 73
  5. Very specific:

    If /fetch-paper returns NOT_FOUND for a source, do not fabricate its content: keep it out of the Review's cited claims and note it under …

    line 89
  6. Very specific:

    Write the review document to docs/reviews/{review-id}.md using the template at .claude/skills/literature-review/templates/review.md.

    line 96
  7. Very specific:

    Read the existing review from docs/reviews/.

    line 110
  8. Distinctive:

    If the edit adds or removes a source, re-run the relevant slice of Phase 2 (relevance check) and Phase 3 (persist) for any new source …

    line 112
  9. Very specific:

    Update the Status field if appropriate (draft → active → superseded).

    line 114
  10. Very specific:

    All reviews live in docs/reviews/

    line 120
  11. Very specific:

    When starting a new review, check if a similar one already exists before searching (Glob docs/reviews/*.md and scan titles/scope lines)

    line 123

“Common” is measured: kept by at least 6 in 10 of the 72 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Purpose

This skill produces a literature review: a scoped survey of what the published record says about a specific topic, backed by persisted sources (docs/papers/, docs/references/) and a synthesis document (docs/reviews/). The deliverable is a review of the literature itself — not a proposed change to this project.

Reuse, don't reimplement: this skill never fetches a PDF or writes a reference doc itself — it always delegates to /fetch-paper and /references, which already own that pipeline (dedup against docs/references/INDEX.md, slug format, extraction agents).


When to invoke

  • User asks to review, survey, or summarize the literature on a topic
  • User wants to view, update, or refer back to an existing review
  • Invoke explicitly: /literature-review or /literature-review edit <review-id>

Flow

Phase 0: Scoping        →  plan mode → interrogate topic scope, meaning, background
Phase 1: Search plan     →  present search terms & criteria → confirm before searching
Phase 2: Broad search    →  search widely → filter to ~10 strictly relevant sources → confirm shortlist
Phase 3: Persist         →  /fetch-paper + /references for each confirmed source
Phase 4: Formalization   →  write review document to docs/reviews/
Phase 5: Edit mode       →  (optional) read existing review → targeted edits

Phase 0 — Scoping (Plan Mode)

Enter plan mode as soon as a topic is named. A vague topic produces a vague search — this phase exists to prevent that.

Ask targeted questions — do not ask for anything inferable from context (e.g. docs/project/research.md, existing proposals, existing reviews):

  1. Meaning: "When you say '{topic}', what specifically do you mean — is there a narrower formulation (a method, a metric, a comparison) you have in mind?"
  2. Background: "What's driving this review — a design decision, a proposal you're weighing, a gap you noticed? This shapes which sources count as 'directly applicable' versus merely adjacent."
  3. Boundaries: "What should the search explicitly exclude? (e.g. a related subfield, a method already covered in an existing review or proposal, an era of work no longer relevant.)"
  4. Comparison axis: "Should sources be organized by sub-theme, by chronology/lineage, or by competing approaches?" — this decides the Review section's structure in Phase 4.

The Background answer is scoping input only — it must never surface in the written document. It shapes which sources get searched for and which count as "directly applicable" (Phase 1/2), but the review itself is a synthesis of the literature, not of this project. See the Objectivity guardrail in Phase 4: the document may not name a proposal, another internal doc, this project, or any project-specific decision it is informing.

Summarize the refined scope in one sentence and ask: "Does this scope match what you want reviewed?"

Checkpoint: User confirms the scope sentence before any search begins.


Phase 1 — Search plan

Before searching, state:

  • The search terms/queries to be used (derived from the confirmed scope, not the raw topic)
  • The relevance criteria from /references (direct relevance, strength of connection, academic standing) as the filter that will bring the candidate pool down to strictly relevant sources
  • The target count: ~10 sources, fewer if the scoped topic is narrow and fewer genuinely qualify — never pad with loosely related sources to hit the number (per /references: "prefer fewer, highly relevant references over a broad list of loosely related ones")

Ask: "Proceed with this search?"

Checkpoint: User approves the search plan before Phase 2 begins.


Phase 2 — Broad search

  1. Search broadly (WebSearch, Semantic Scholar/OpenAlex queries as used by /fetch-paper's resolution phase) using the confirmed terms. Cast wider than 10 — the goal is a candidate pool to filter down, not a first-10-hits list.
  2. Apply the /references relevance criteria to every candidate. Discard anything that is only superficially related or whose connection to the scoped topic requires a stretch.
  3. The ~10-source count is exclusive to sources that directly solve, apply to, or are demonstrated on the scoped topic itself — not sources that merely discuss an adjacent mechanism, technique, or theory in general terms. A general survey/method paper that never touches the scoped topic concretely (e.g., a curriculum-learning paper tested only on an unrelated domain, when the topic is a specific environment) does not count toward the 10, even if it is genuinely good supporting context. Such papers may still be cited in the Review section where they provide real support for a specific claim, and may still appear in the References table alongside the 10 — they just aren't counted against the target when judging whether the shortlist is "exclusive enough." When in doubt whether a candidate is direct or merely adjacent, ask the user rather than silently including or excluding it.
  4. Check docs/references/INDEX.md for candidates already covered — keep them in the shortlist (they still count toward the 10) but they will skip Phase 3's fetch step.
  5. Present the shortlist — title, authors, year, one-line reason for inclusion, and whether each is a direct-topic source (counted toward the 10) or a supporting/adjacent source (not counted) — and ask: "Confirm this source list, or should any be swapped out?"

Checkpoint: User confirms the final source list before any fetching happens.


Phase 3 — Persist sources

For each confirmed source not already in docs/references/INDEX.md:

  1. Invoke /fetch-paper {title or arXiv ID or DOI} — this downloads the PDF to docs/papers/{slug}.pdf and can bootstrap the reference skeleton.
  2. Invoke /references to run the extraction pipeline and produce the full docs/references/{slug}.md, including its INDEX.md row.

For sources already indexed, skip straight to using their existing docs/references/{slug}.md in Phase 4 — do not re-fetch or re-extract.

If /fetch-paper returns NOT_FOUND for a source, do not fabricate its content: keep it out of the Review's cited claims and note it under Gaps and open questions in Phase 4 as a source that could not be retrieved.


Phase 4 — Formalization

  1. Generate a unique review ID: YYYY-MM-DD-{topic-slug} (e.g. 2026-07-14-quality-diversity-curricula). The slug is 3–5 words in kebab-case describing the reviewed topic.
  2. Write the review document to docs/reviews/{review-id}.md using the template at .claude/skills/literature-review/templates/review.md.
  3. Structure the Review section per the comparison axis agreed in Phase 0 (sub-theme / chronology / competing approaches).
  4. Objectivity guardrail: the entire document is a synthesis of the literature and nothing else. It does not synthesize what this project should do about the topic, and it does not mention this project, any other internal document (a proposal, docs/project/research.md, another review, a fix doc), or the design decision/gap that motivated the review. The Background section states why the topic matters in field terms (the problem, question, or gap in the literature itself) — never "this is being reviewed to check proposal X" or "this informs decision Y." If the scoping conversation named a specific internal document or decision, that context stays in the conversation and shapes source selection only; it must not appear in docs/reviews/{review-id}.md. Recommendations for this project, or explicit comparisons against a specific internal document, belong in a separate proposal or in conversation — never in the review document.
  5. Every claim in the Review section must trace to a specific source in the References table — cite inline per .claude/rules/references.md style (Author et al., Year). Claims that can't be attributed to a specific source get removed or moved to Gaps and open questions as an explicit open question.
  6. The References table must list exactly the sources confirmed in Phase 2 — no additions, no silent drops.

Announce the file path to the user after writing: Review written to docs/reviews/{review-id}.md.


Phase 5 — Edit mode

Triggered when the user references an existing review by ID, filename, or topic.

  1. Read the existing review from docs/reviews/.
  2. Identify which section(s) the user wants to change.
  3. If the edit adds or removes a source, re-run the relevant slice of Phase 2 (relevance check) and Phase 3 (persist) for any new source before editing the Review or References sections.
  4. Apply edits using targeted replacements — do not rewrite the whole document.
  5. Update the Status field if appropriate (draft → active → superseded). No other bookkeeping is required.

Naming and indexing

  • All reviews live in docs/reviews/
  • File name: {YYYY-MM-DD}-{topic-slug}.md
  • Status field values: draft, active, superseded
  • When starting a new review, check if a similar one already exists before searching (Glob docs/reviews/*.md and scan titles/scope lines)

Synergy with other skills

SkillWhen to invoke
/fetch-paperPhase 3, once per confirmed source not already saved
/referencesPhase 2 (relevance criteria), Phase 3 (extraction pipeline), Phase 5 edits that touch sources

References

#TitleAuthorsYearVenueRelevance
1Preferred Reporting Items for Systematic Reviews and Meta-Analyses: The PRISMA StatementMoher, D.; Liberati, A.; Tetzlaff, J.; Altman, D.G.2009PLOS Medicine 6(7): e1000097. DOI: 10.1371/journal.pmed.1000097Establishes the identify-screen-eligibility-include funnel that structures Phases 2–3 (broad candidate pool filtered to a confirmed, criteria-justified source list)
2Guidelines for Conducting Systematic Literature Reviews in Software EngineeringKitchenham, B.; Charters, S.2007EBSE Technical Report EBSE-2007-01Grounds the scoping-before-searching order (define the research question and inclusion/exclusion criteria before the search) that motivates Phase 0 preceding Phase 2

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.