-
/oss:research-repo·$research-repoSet up a research project's folders the same way every time, with a sources spine and a
CLAUDE.mdfile the AI reads first. Or point it at an existing project to check that the layout already holds.
-
/oss:sitrep·$sitrepFind out where a project stands after time away. Reads whatever handoff and log files the repo keeps, checks what git actually shows, and says plainly where the two disagree.
-
/oss:finished·$finishedClose out a work session by writing down what changed, so the next one doesn't start from scratch. Separates what was verified from what was only attempted.
-
/oss:advisor·$advisorAsk for a second opinion before you commit to a decision or call work done. In Claude Code your own session holds the main seat, on Opus or on Sonnet for cheaper sustained work, and the advisor seat is always Fable, pinned to max reasoning effort no matter what your session runs at. In Codex it is always GPT-6 Astra at xhigh, the flagship tier at its top effort setting, because the point of asking is a stronger reviewer rather than a cheap one. It is an escalation from a working model, not a peer check: a Fable session should reach for orchestrate's Astra peer or the committee instead. Advice only. It never edits your files.
-
/oss:orchestrate/oss:fable-orchestrate/oss:opus-orchestrate$orchestrateRun a hard task past several AI models at once. The skill reads which model your session is on and takes the matching lead role. In Claude Code, Fable keeps the hard thinking itself and hands off mechanical work to Sonnet, wide parallel reasoning to Opus subagents, and selected outside checks to GPT-6 Astra; Opus is itself the deep reasoner and delegates to fan out. The two older Claude names still work and force the lead. In Codex, an active GPT-6 Astra or GPT-5.6 Sol session can lead: Astra keeps compact hard reasoning in-session, while Sol escalates unusually difficult units to Astra. Both modes route bounded and mechanical work to lower GPT-5.6 tiers and can use a Claude peer.
-
/oss:spawn·$spawnLaunch full, independent AI sessions to work on separate tasks at the same time — not quick helpers, but complete sessions you can watch and steer in their own window, each in its own copy of your project so nothing collides. Picks the strongest way to do this that's available (a terminal manager called herdr, then tmux, then a plain background process) and merges each one's finished work back in once it passes.
-
/oss:diverge/oss:diverge-codex·$divergeBefore committing to one approach, generate three to five different ones, from the safe option to the deliberately unconventional, so you're choosing between real options instead of the first idea. Runs entirely in whatever model and effort your current session already has, with no separate model call. With
--codex(or the older diverge-codex name) Codex generates the alternatives and, once you choose, implements the chosen one. -
/oss:research-grill$research-grillGet interviewed about a research idea, design, or draft until nothing is left silently assumed. Each round asks every question that can be asked now, numbered, in plain language, with a one-line "why this matters" and a recommended answer; facts get looked up rather than asked; every settled decision is written to a file. Three stages: idea to falsifiable question, question to full design, design or draft to the objections a reviewer would raise. Works for a BA thesis or a grant. Adapted from Matt Pocock's grill-me.
-
/oss:research-wayfinder·$research-wayfinderPlans a study across many sessions instead of one long conversation, so the big open decisions in a design — what you're actually estimating, how you'll identify it, whether you have the statistical power, how you'll measure it — get resolved one at a time and written down, instead of re-argued every time you sit down. Ends with a pre-analysis plan ready to write up. Adapted from Matt Pocock's
wayfinder, a similar tool built for planning software projects. -
/oss:conjoint-design·$conjoint-designWork out which attributes to test and how many respondents you need, then how you'll estimate each attribute's effect (AMCE/AMIE), all before you run the experiment.
-
/oss:conjoint-diagnostics·$conjoint-diagnosticsRun through a checklist of what typically goes wrong in a conjoint experiment, from the design through the analysis, and catch it before submission.
-
/oss:conjoint-cleaning·$conjoint-cleaningTurn a raw conjoint survey export from Qualtrics into a clean, analysis-ready dataset.
-
/oss:survey-design·$survey-designWrite survey questions and response scales that measure what you intend and flow logically, without quietly biasing respondents toward a socially acceptable answer.
-
/oss:qualtrics-ops/oss:survey-flow-audit·$qualtrics-opsMake changes to a live Qualtrics survey through the API without breaking what respondents see — publishing, quotas, flow routing, and vendor redirects, each verified by reading the result back. Its
auditmode (or the older survey-flow-audit name) is the read-only pre-fielding check: consent before anything, force-response completeness, quotas, vendor redirects, anti-bot instrumentation, language-arm symmetry, with an optional browser walk. -
/oss:survey-data-audit·$survey-data-auditCheck fielded survey data before anyone touches the outcomes — were bots screened, did every registered field arrive, does the sample match the quotas — and produce the quality appendix for the paper.
-
/oss:cross-national-design·$cross-national-designAdapt a survey experiment to run fairly across multiple countries, with enough statistical power in each and wording localized rather than translated word-for-word. Includes a check for hidden bias between countries.
-
/oss:list-experiment·$list-experimentDesign and check a list experiment (the item-count technique) — a way of asking about sensitive topics without anyone naming their own answer directly.
-
/oss:topic-modeling·$topic-modelingSet up a structural topic model, which groups documents into themes and shows how those themes relate to variables you care about, with the diagnostics to know whether it worked.
-
/oss:text-classification·$text-classificationHave an AI model sort text into categories you define, following a written codebook, then check its accuracy the way you'd check a human coder's.
-
/oss:model-council-voting·$model-council-votingHave several AI models independently label the same data, then combine their votes and measure how much they agreed — inter-rater reliability, but with models instead of people.
-
/oss:model-committee·/oss:model-committee-astra·/oss:model-committee-opus·/oss:model-committee-fable·/oss:model-committee-sol·$model-committeeHave two AI models (GPT-6 Astra and Claude Opus) discuss a judgment call with each other before settling on an answer, instead of asking just one. The chair is a parameter, not a separate skill, and it is a premier model that never sits on the committee:
/oss:model-committeehas Fable run the meeting,/oss:model-committee-astrahands the gavel to GPT-6 Astra — in which case the GPT debater drops to the Sol tier so the chair is never grading its own family's twin —/oss:model-committee-opuskeeps the cheap in-session Opus chair, and/oss:model-committee-solis the legacy GPT-5.6 chair with a Terra debater. -
/oss:llm-calibration-logprobs·$llm-calibration-logprobsCheck how confident an AI model actually is in each answer, using the model's own token probabilities rather than just asking it to rate its confidence.
-
/oss:doc-to-markdown·$doc-to-markdownHand the model a PDF, Word file, or slide deck and get a sensible read of it — deciding when the file can simply be read, when it needs converting, which converter suits that particular document, and whether the resulting text is worth keeping.
-
/oss:vlm-ocr/oss:vlm-ocr-evaluation/oss:vlm-ocr-pipeline/oss:post-ocr-cleanup$vlm-ocrTurn scanned documents into text in three phases. Evaluate compares OCR systems on a stratified ground-truth sample with error rates per language and script, so you pick a model on evidence before a bulk run; run builds the vision-language-model pipeline (image handling, prompts, batching, provenance); clean corrects the output with model and rule-based passes, quality diagnostics, and multilingual handling. The three older names still work and force a phase.
-
/oss:hypothesis-building·$hypothesis-buildingTurn a research idea into a hypothesis that could actually be proven wrong — with a causal diagram and a plan for testing it, including the possibility of no effect.
-
/oss:literature-review·$literature-reviewMap out what's already been studied on a topic and spot the real gaps, then plan how to synthesize the literature you've found.
-
/oss:narrative-building·$narrative-buildingDraft or tighten a paper's introduction so the argument actually builds — why the question matters, and what follows if the answer goes one way or the other.
-
/oss:pre-registration-writing·$pre-registration-writingWrite a pre-analysis plan that states what you'll measure and how you'll analyze it, registered before you see the results.
-
/oss:methods-reporting·$methods-reportingCheck that your methods section reports everything expected by standard research-reporting checklists (CONSORT, JARS, DA-RT).
-
/oss:paper-tex·$paper-texTake a draft in any format, Word or Markdown or plain text, and typeset it into a properly formatted LaTeX paper, ready for a specific journal's requirements.
-
/oss:figures·$figuresDesign a publication-quality figure, choosing chart type, color, scales, and captions to make the data as easy to read as possible.
-
/oss:tables·$tablesDesign a publication-quality table, from column order and row grouping to number formatting and what the notes underneath should say.
-
/oss:citation-check·$citation-checkCheck every citation in a manuscript against its reference list, catching ones that don't exist, don't match, or are formatted wrong.
-
/oss:fact-check·$fact-checkCheck whether the sources a manuscript cites actually say what the manuscript claims they say.
-
/oss:figure-table-audit·$figure-table-auditA final pass over a manuscript's figures and tables before submission, checking captions and cross-references along with the statistical notes.
-
/oss:replication-package/oss:fair-check·$replication-packageBuild or check a replication package, the folder of data and code plus the documentation someone else would need to reproduce your results. Its
auditmode (or the older fair-check name) checks a finished package and manuscript, including a FAIR block: data, code, materials, and prompts available under stated licenses and persistent identifiers, with reuse conditions spelled out.
-
/oss:paper-review-lite/oss:paper-review-lite-codex·$paper-review-liteRun a manuscript through a pre-submission review that checks its claims against actual quotes from the text, catching overclaiming before a real reviewer does. With
--codex(or the older paper-review-lite-codex name) the same audit runs on Claude and on Codex independently and the findings are cross-checked with a confidence column. -
/oss:presubmitA guided walkthrough of a much larger, 30-plus-stage automated pre-submission review tool (the standalone presubmit CLI, API-billed). It routes each stage to one of four Claude tiers by task difficulty. Haiku handles mechanical extraction, Sonnet handles bounded verification, Opus runs the adversarial Red Team, and Fable handles the Blue Team/Assessment/Reviewer synthesis that judges the Red Team's findings — so the critics and their judge aren't sharing a model. ~$2–4 per full run on an 80-page paper. A short article runs well under $2.
-
/oss:journal-review·$journal-reviewDraft a referee report on someone else's manuscript, as if you were reviewing it for a journal.
-
/oss:referee-response$referee-responseOrganize your response to peer review without letting the AI write the science. It pulls every distinct point out of the referee reports and the editor's letter, tags severity and type, orders the revision so upstream changes come first, flags the points you might reasonably push back on as questions for you, and builds the response letter as a numbered comment / response / location table with the substantive answers left as placeholders for you to fill. A check mode then confirms every point is answered and every cited location exists.
No skills match your search. Try a different word, or clear the category filter.
This index mirrors Open Science Skills as of 17 September 2026 (v2.32.0). It includes 40 skills on Claude Code (/oss:name) and 39 on Codex ($name). Most run on both platforms, with a few platform-specific exceptions. The library also carries an experimental deliverable pipeline, not listed here: it manages writing projects rather than research methods, and its interfaces are still moving.
Eleven skills load automatically when relevant. The other 27 load only when invoked by name, so they use no session resources until needed. All aliases still work. The Codex library follows the same logic.