Skills

Skills are the unit of work. Each one is a directory holding a SKILL.md: frontmatter with a name and a description full of trigger phrases, then the workflow the model follows. There are 35 of them, in eight categories.

You do not have to invoke a skill by name. The description is what the model matches your phrasing against, so “find recent papers on X” reaches literature-search on its own. The trigger column below is a hint, not a magic word.

The same 35 skills install into OpenAI Codex as researcher-<name> (see Codex), where you can also call one explicitly, for example $researcher-literature-search.

Research & Discovery (8)

Skill	What it does	Typical trigger
literature-search	Multi-source search across OpenAlex, Crossref, Semantic Scholar, arXiv, PubMed, and Scite. Deduplicates by DOI then title, ranks results, and in systematic mode records a PRISMA flow and a search-provenance record.	”Do a systematic literature search on X”
research-convergence	Deep research that converges on an argument rather than a pile of PDFs. Three modes: Socratic, Full Research, Flash.	”Help me build the thesis for this”
fact-checking	Verifies a claim against the literature and returns a verdict (Supported, Contested, Partially Supported, Contradicted, Unsupported) with the evidence behind it. Returns Unsupported rather than invent a source.	/researcher:fact-check <claim>
sota-finder	Finds state-of-the-art results for a benchmark, with performance timelines and comparison tables. Numbers are traced to the paper that reported them.	/researcher:sota <benchmark>
research-gaps	Identifies methodological, empirical, and theoretical gaps in a body of literature, ranked by impact potential.	”What is missing in this literature?”
citation-context	Classifies citations as supporting, contrasting, or mentioning, and audits a manuscript for sources it misrepresents.	”Check how I framed these citations”
citation-audit	Audits every citation in a manuscript for existence (does it resolve, is it retracted) and faithfulness (does the source support the claim), refusing a clean verdict on any refusal-grade finding.	/researcher:verify-citations
research-pipeline	Drives a manuscript through Plan, Retrieve, Synthesize, Draft, Review, Compile, and Format, with a checkpoint at every stage and a compile gate before formatting.	”Run the pipeline from question to draft”

Full walkthrough: Research & Discovery guide.

Systematic Review (4)

Skill	What it does	Typical trigger
systematic-review	The end-to-end workflow over the event ledger: lock a protocol, run per-database strategies, dedup, dual screen with blinded adjudication, then hand off to extraction, appraisal, synthesis, and reporting. Refuses to screen before a protocol lock. Verifies every included reference on identity and flags an inconclusive result for human review rather than dropping it. PRISMA 2020 is derived from the ledger, never stored.	/researcher:systematic-review
extraction-tables	Elicit-style structured extraction: per-paper values anchored to full-text passages, with the verification layer stated on every cell, “not reported” instead of an invented value, and typed effect-size columns the meta-analysis consumes mechanically.	”Build an extraction table from these papers”
contradiction-detection	A claim-by-source contradiction matrix, each cell carrying its quote and verification layer, feeding GRADE’s inconsistency domain with concrete cited disagreements.	”Find evidence that contradicts this claim”
literature-monitoring	Living reviews: saved verbatim searches, a diff of new records against the seen list on rerun, feeding a fresh screening batch under the same locked protocol.	/researcher:watch-topic

The systematic-review vertical is built on the event ledger, so the PRISMA 2020 flow and checklist are recomputed from events rather than stored, and protocol locking, dual screening, and appraisal (RoB 2, GRADE) all leave an auditable trail. New in 0.5.0.

Planning & Design (4)

Skill	What it does	Typical trigger
brainstorming	Socratic refinement, from a vague hunch to a precise research question, hypotheses, and a method.	/researcher:brainstorm
experiment-design	Designs studies with controls, sample sizing, power analysis, ablations, and a reproducibility checklist.	/researcher:design-experiment <question>
statistical-analysis	Test selection, assumption checking, generated analysis code (Python, R, MATLAB), and APA-formatted reporting.	”Which statistical test should I use?”
manuscript-setup	Scaffolds the project: per-section files, bibliography, figures and tables folders, config.yaml, terminology tracking. LaTeX, Word, or both.	/researcher:new-manuscript

Full walkthrough: Planning & Design guide.

Writing & Revision (6)

Skill	What it does	Typical trigger
paper-drafting	Outline mode and section-by-section drafting, with cross-referencing, consistent terminology, and journal-aware formatting.	/researcher:draft-section <section>
writing-style-analysis	Reads your past papers and produces a reusable style-profile.yaml that drafting then writes to.	”Analyze my writing style from these papers”
peer-review	Five reviewer personas plus an Editor-in-Chief, with a scoring rubric and a decision. External model reviewers (OpenAI, Gemini, Ollama) are a documented integration point, planned and not implemented.	/researcher:review-paper
revision-management	Parses reviewer comments into a tracked roadmap and generates LaTeX tracked changes (the changes package or latexdiff). Word tracked changes are planned, not implemented.	/researcher:revise <round>
response-to-reviewers	Point-by-point response document: every comment answered, changes quoted, disagreements argued with evidence.	”Write the response to reviewers”
cover-letter	Journal-appropriate cover letter, with the tone matched to the journal’s tier.	”Write a cover letter for this journal”

Full walkthrough: Writing & Revision guide.

Visualization & Figures (6)

Skill	What it does	Typical trigger
visualization	Publication-quality plots as runnable code: matplotlib, seaborn, ggplot2, plotly, pgfplots. Chart-type selection and statistical annotation included.	”Plot these results”
tikz-diagrams	Architectures, flowcharts, state machines, timelines, and pipelines in TikZ/PGF.	”Draw the system architecture”
plotneuralnet	Neural network architecture diagrams with 3D layered boxes, as a self-contained single-file .tex. Presets for VGG, ResNet, U-Net, Transformer.	”Make a CNN architecture diagram”
figure-suggestions	Reads the manuscript and recommends which figures it needs, of what type, and where they go.	”What figures should this paper have?”
latex-tables	Booktabs tables from CSV or JSON: significance markers, bold best results, multi-column, landscape, longtable.	”Turn this CSV into a results table”
image-prompt-crafting	Prompts for external image generators, for conceptual illustrations, graphical abstracts, and cover art only. Never for data or results figures, and always with an AI-disclosure caption.	”Draft a graphical abstract prompt”

All the figure-producing skills share named style presets (default, Nature, IEEE). A preset restyles typography, sizing, and color; it never touches your data. Full walkthrough: Visualization & Figures guide.

Publishing & Formatting (3)

Skill	What it does	Typical trigger
journal-finder	Ranked, reasoned journal recommendations, filtered by impact factor, quartile, indexing, open access, and APC.	/researcher:find-journal
conference-finder	Conferences and workshops with deadlines, rankings, acceptance rates, and locations.	/researcher:find-conference
journal-formatting	Applies a target journal’s requirements from a local database of 16 publisher and journal profiles; anything else is looked up from the publisher’s author guidelines. Validates compliance.	”Format this for IEEE Access”

Full walkthrough: Publishing & Formatting guide.

Code & Implementation (2)

Skill	What it does	Typical trigger
implementation	Experiment scripts, data pipelines, and evaluation code, with reproducibility (seeds, environment, config) enforced.	”Implement this training loop”
code-analysis	Reads your codebase and drafts the methods section from it: pseudocode, complexity analysis, algorithm environments.	”Write methods from this repo”

Both carry context: fork and agent: code-agent in their frontmatter, so they run in the Sonnet-pinned Code Agent and your session’s higher-tier budget stays on research and writing. They are the only two skills that fork. Full walkthrough: Code & Implementation guide.

Output & Management (2)

Skill	What it does	Typical trigger
word-output	DOCX generation through templates/word/build-docx.js (node, built on the docx library): title page, numbered headings, paragraphs, lists. Tracked changes, comments, and table emission are specified but not implemented yet.	”Give me this in Word”
citation-management	Maintains library.bib: import, DOI validation against Crossref, retraction detection, predatory-venue flags, format conversion, and a citation-integrity audit. Zotero through the user-installed zotero-mcp server; Mendeley by manual export.	”Validate my bibliography”

Both are covered in the Publishing & Formatting guide.

Where to go next

• The Commands that route into these skills.
• The Agents that orchestrate several of them at once.
• The Connectors the research skills call out to.