Open Science Skills

Claude Code OpenAI Codex version license Claude skills Codex skills updated sources PRs welcome

Open Science Skills is a library of 38 agentic skills for Claude Code, with a parallel 37-skill library for OpenAI Codex, written for computational social scientists and digital humanists. Each skill is meant to work the way the field expects. Identify the data-generating process before proposing an estimator, and design experiments and instruments to a standard. Drafts are held to established reporting norms.

The library follows the research lifecycle. It covers survey design, list experiments, topic modeling, LLM text classification, VLM-based OCR pipelines, manuscript QA, multi-model orchestration, and transparent reporting under APSA, JARS, DA-RT, TOP, and FAIR expectations. Every skill is grounded in published methods sources and based on best practices for writing skills. See SOURCES.md for the bibliography of 150+ works consulted.

This is the toolkit I use in my own research, and it grows as I add sources and skills. The authoring is mine, with editing help from Opus 5, Fable 5.1, and ChatGPT 5.5/6.

Platform Skills Invoke
Claude Code 38, as the oss plugin /oss:skill-name
OpenAI Codex 37, as the codex/ library $skill-name

The two libraries differ only in invocation and tooling. The Codex side omits presubmit; its orchestrate is the Codex-native version led by the active GPT-6 Astra or GPT-5.6 Sol session, with mode-aware routing between them. See codex/README.md.

Model selection and reusable migration practices are documented in MODEL-POLICY.md.

Quick start · Skills · Recommended · How skills trigger · Installation · Sources · Contributing · License

Quick start

Install the plugin from the marketplace, user-wide across all projects, or add --scope project for one project only:

# Step 1: Register the marketplace (one-time)
claude plugin marketplace add scdenney/open-science-skills

# Step 2: Install the plugin
claude plugin install oss@open-science-skills

# Project-only install
claude plugin install oss@open-science-skills --scope project

Then invoke a skill explicitly, for example /oss:conjoint-design, or just describe your task in plain language and let the matching skill load on its own — for the auto-triggering core; on-demand skills load only by name (see On-demand skills under Skills).

On Codex there is no plugin. Install the skills library instead (see Codex).

Skills

On-demand skills. Eleven skills are the auto-triggering core — Claude loads them from context when the task matches: citation-check, doc-to-markdown, fact-check, figures, literature-review, paper-review-lite, qualtrics-ops, referee-response, replication-package, research-repo, spawn. The other 27 are on-demand: they cost nothing in a session until you invoke them by name (/oss:conjoint-design), and Claude does not suggest them unprompted. That keeps a research session's always-on cost near 3k tokens instead of 10k. The orchestration and deliberation skills (orchestrate, advisor, model-committee, model-council-voting, diverge, journal-review) are on-demand by design, since they start subagents or external models; the rest are specialists that fire rarely enough that a name is the better trigger. Every skill still resolves as a slash command and through its aliases. The Codex library applies the same rule with its own evidence (allow_implicit_invocation: false; see codex/README.md).

Skills are grouped by where they fall in a project. Unless the Platform column says otherwise, a skill runs on both Claude Code (/oss:name) and Codex ($name). Names retired in v2.25.0 (fable-orchestrate, opus-orchestrate, diverge-codex, paper-review-lite-codex, survey-flow-audit, vlm-ocr-evaluation, vlm-ocr-pipeline, post-ocr-cleanup, fair-check) still work as aliases that call the merged skill with its mode set.

Project Setup

Skill Platform Command What it does
research-repo Both /oss:research-repo Scaffold a new research project around its source library, or audit an existing one. The sources folder is the spine. From there it builds the references file and the intake tooling, plus the analysis and manuscript folders and a place for reviews.

Workflow & Orchestration

Skill Platform Command What it does
orchestrate Both /oss:orchestrate · /oss:fable-orchestrate · /oss:opus-orchestrate · $orchestrate Run a multi-model workflow. Claude reads the session model and takes the matching Fable 5.1 or Opus 5 lead role. Codex likewise detects the active GPT-6 Astra or GPT-5.6 Sol session: Astra keeps compact hard reasoning in-lead, while Sol escalates unusually difficult units to Astra. Both Codex modes route bounded and mechanical work to lower GPT-5.6 tiers and can use a Claude peer.
spawn Both /oss:spawn Spawn full peer sessions in new terminal panes — real sessions, not subagents — each in its own git worktree on a directed task with a contract brief. Detects herdr, tmux, or a plain terminal and takes the strongest path; the lead monitors without babysitting and merges each branch back.
advisor Both /oss:advisor / $advisor Escalate one decision point from a working model to an independent second reviewer before committing to an interpretation or calling a task done. Your session is the main seat, on Opus 5 or on Sonnet 5 for cheaper sustained work; the advisor seat is Fable 5.1, pinned to max reasoning effort. Not for a Fable session — a second Fable is not a check; a Fable lead goes cross-vendor through orchestrate's Astra peer or the committee instead. The Codex counterpart escalates to Astra/xhigh the same way.

Ideation

Skill Platform Command What it does
diverge Both /oss:diverge · /oss:diverge-codex Before implementing, generate three to five distinct approaches labeled by how they differ, then pause for you to choose. --codex (or the alias) has Codex (GPT-6 Astra at xhigh) generate and, once chosen, implement.

Research Design

Skill Platform Command What it does
research-grill Both /oss:research-grill Interview the researcher in rounds until nothing is silently assumed: idea → falsifiable question, question → full design, design or draft → the reviewer's objections. Plain-language questions with a recommended answer each, facts fetched rather than asked, decisions written to a wayfinder ticket or decisions.md. Works for a BA thesis or a grant.
research-wayfinder Both /oss:research-wayfinder Plan a study as a decision map that outlives any single session — typed tickets for estimand, identification, power, and measurement, resolved one per session until the design is pre-registerable. Adapted from Matt Pocock's wayfinder.
conjoint-design Both /oss:conjoint-design Design conjoint experiments, including the attribute architecture and the statistical power that architecture implies. Covers AMCE and AMIE estimation.
conjoint-diagnostics Both /oss:conjoint-diagnostics Check a conjoint design and its analysis for integrity, measurement error, external validity, and sound interpretation.
conjoint-cleaning Both /oss:conjoint-cleaning Reshape a Qualtrics conjoint export into analysis-ready long format, with choice mapping, translation, pilot detection, and validation.
survey-design Both /oss:survey-design Write survey instruments. Covers question wording, scales, flow, pretesting, respondent burden, and social-desirability mitigation.
qualtrics-ops Both /oss:qualtrics-ops · /oss:survey-flow-audit Operate a live Qualtrics survey via the v3 APIs without breaking fielding: publish gating, quotas, flow routing, embedded data, panel-vendor redirects, read-back verification. audit (or the alias) is the read-only pre-fielding audit: consent gates, force-response completeness, quotas, redirects, anti-bot instrumentation, language-arm symmetry, optional browser walk.
survey-data-audit Both /oss:survey-data-audit Audit fielded survey response data for registered elements, data quality, bot and AI-automation screening, and sample integrity. Emits an appendix-ready quality report.
cross-national-design Both /oss:cross-national-design Design survey experiments that run across countries, with per-country power and measurement equivalence checks. Includes instrument localization.
list-experiment Both /oss:list-experiment Design and diagnose list experiments (the item count technique), from sensitivity assessment through estimation and placebo checks.

Analysis

Skill Platform Command What it does
topic-modeling Both /oss:topic-modeling Fit structural topic models, choosing the topic count by coherence and exclusivity rather than by eye. Covers covariate specification and what to report.
text-classification Both /oss:text-classification Classify text with LLMs. Covers codebook design, human-in-the-loop workflows, validation, and agreement statistics.
model-council-voting Both /oss:model-council-voting Use a panel of models as independent coders under a consensus rule stated in advance, then read their disagreement with chance-corrected agreement statistics (the kappa and alpha families). Includes checks for correlated errors across jurors.
model-committee Both /oss:model-committee, /oss:model-committee-astra, /oss:model-committee-opus, /oss:model-committee-fable, /oss:model-committee-sol Have GPT-6 Astra and Claude Opus 5 deliberate toward one decision. They propose independently, critique each other, revise, and converge under a rule fixed before they start. The chair is a premier model that is never a member: Fable 5.1 by default, or GPT-6 Astra (-astra), in which case the GPT member steps down to Sol. -opus keeps the cheap in-session Opus chair (which is also a member), -fable aliases the default, and -sol is the legacy GPT-5.6 chair with Terra as its member.
llm-calibration-logprobs Both /oss:llm-calibration-logprobs Turn token log-probabilities into per-decision confidence, then measure calibration against human labels (ECE and Brier scores, plus reliability diagrams).

Corpus Processing

Skill Platform Command What it does
vlm-ocr Both /oss:vlm-ocr · /oss:vlm-ocr-evaluation · /oss:vlm-ocr-pipeline · /oss:post-ocr-cleanup OCR for scanned material in three phases. evaluate compares OCR systems on stratified ground truth with CER/WER per language and script; run builds the vision-language-model pipeline (model choice, image handling, prompts, batching, provenance); clean corrects the output with LLM and rule-based passes, quality diagnostics, and multilingual handling. The three aliases force a phase.
doc-to-markdown Both /oss:doc-to-markdown Read or convert any document a research workflow hands you. Decides whether to read the file directly or convert it, routes to the right converter for the document's actual structure, and decides whether the Markdown is a tracked artifact or a scratch file.

Writing & Reporting

Skill Platform Command What it does
hypothesis-building Both /oss:hypothesis-building Turn a research question into falsifiable causal hypotheses using DAGs, counterfactuals, equivalence testing, and a stated smallest effect size of interest.
literature-review Both /oss:literature-review Build or audit a literature review. Produces an evidence map, an assessment of the closest prior work, gap verdicts, and a synthesis plan.
narrative-building Both /oss:narrative-building Draft or audit a paper's introduction so it gets from the "why" to the "if-then," and keep multi-experiment papers coherent.
pre-registration-writing Both /oss:pre-registration-writing Write a pre-analysis plan. Covers structure, registry choice, analysis strategy, and how to document any deviations.
methods-reporting Both /oss:methods-reporting Check a methods section against a 40-item reporting checklist drawn from CONSORT and JARS, plus the DA-RT transparency standards.
paper-tex Both /oss:paper-tex Typeset a draft as house-style LaTeX from Markdown, Word, or other formats. Builds the PDF and prepares it for a specific journal.

Figures & Tables

Skill Platform Command What it does
figures Both /oss:figures Design publication-quality figures. Covers chart choice, scales, color, legend order, self-contained captions, and reproducible code.
tables Both /oss:tables Design publication-quality tables. Covers column order, row grouping, precision and uncertainty, self-contained notes, and reproducible code.

Manuscript QA

Skill Platform Command What it does
citation-check Both /oss:citation-check Check citations for in-text and reference parity, then for working DOIs and fabrication risk (via Crossref and OpenAlex). Also checks citation style.
fact-check Both /oss:fact-check Verify that each in-text claim is actually supported by its cited source, reading the source's Markdown in the project's knowledge base. Runs citation-check first.
figure-table-audit Both /oss:figure-table-audit Audit the finished figure and table set for cross-references, text consistency, accessibility, and links to supplementary and replication materials.
replication-package Both /oss:replication-package · /oss:fair-check Scaffold or audit a replication package. Scaffold generates the folder structure, README, master script, figure/table crosswalk, codebook, license, and pre-release checklist. audit (or the alias) 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.

Review & Submission

Skill Platform Command What it does
paper-review-lite Both /oss:paper-review-lite · /oss:paper-review-lite-codex Run a pre-submission self-audit of your own manuscript across argument, numbers, references, writing, figures, and replication. --codex (or the alias) runs the same audit on Claude and Codex (GPT-6 Astra) independently and cross-checks the findings with a confidence column.
presubmit Claude Code /oss:presubmit Set up and run the standalone presubmit CLI, a heavier 30-plus-stage adversarial review pipeline driven by the Anthropic API.
referee-response Both /oss:referee-response Organize and format your response to reviewers: extract every referee point with severity and type, order the revision by dependency, flag defensible pushbacks as questions, and build the response letter with the substantive answers left to you. Never writes the science.
journal-review Both /oss:journal-review Draft a senior referee report on someone else's manuscript, using parallel finder agents and a chief-reviewer synthesis to produce a structured report.

Recommended companion skills

Third-party skills this library recommends and builds on — credited, not claimed, and not counted in the badges. From Matt Pocock's skills (MIT): grill-me (a frontier-rounds design interview; the seed of research-grill, and pairs with diverge), wayfinder (decision-map planning for software work; the source concept for research-wayfinder), and handoff / claude-handoff (handoff documents for a successor session; the seed of spawn). See RECOMMENDED.md for the full write-up and third-party/mattpocock/ for pinned, unmodified reference copies.

How skills trigger

Most skills load on their own. When your prompt matches a skill's description, Claude Code or Codex reads that skill into context and follows it, so you usually don't need to name anything. You can also invoke any skill explicitly, with /oss:skill-name in Claude Code or $skill-name in Codex.

The orchestration and delegated-review skills (orchestrate and its lead aliases, spawn, advisor, model-committee and its chair variants, diverge --codex, and paper-review-lite --codex) run only when invoked explicitly and are on-demand in the catalog (see On-demand skills above), because they start subagents, full peer sessions, or an external model.

Installation

Claude Code

The recommended install is the plugin, shown in Quick start. It registers the marketplace and installs all 38 skills, their slash commands, and the alias commands for retired names. The command prefix is oss:, for open science skills. The marketplace and the repository are both named open-science-skills.

To try the plugin for one session without installing:

git clone https://github.com/scdenney/open-science-skills.git
cd open-science-skills && claude --plugin-dir ./plugin

Clone the repository and run the interactive installer, which lists the skills and installs your choices to ./.claude/skills/ (current project) by default:

git clone https://github.com/scdenney/open-science-skills.git
cd open-science-skills
bash plugin/scripts/install.sh

Other targets and non-interactive selection:

# Install to user-wide skills directory (all projects)
bash plugin/scripts/install.sh --target ~/.claude/skills

# Install specific skills non-interactively
bash plugin/scripts/install.sh --skill conjoint-design survey-design list-experiment

# Install all skills
bash plugin/scripts/install.sh --all --target ~/.claude/skills

Restart Claude Code after installing.

Copy the whole skill folder, since many skills ship reference, asset, or script files their SKILL.md points at (replace your-project with your project's path):

git clone https://github.com/scdenney/open-science-skills.git

# Project-level (current project only) — copy the whole skill folder:
# many skills ship reference/, assets/, or scripts/ files their SKILL.md points at
mkdir -p your-project/.claude/skills
cp -R open-science-skills/plugin/skills/conjoint-design \
   your-project/.claude/skills/

# User-wide (all projects)
mkdir -p ~/.claude/skills
cp -R open-science-skills/plugin/skills/list-experiment ~/.claude/skills/

Manual copy gives auto-trigger only. Slash commands require the plugin.

Codex

Codex discovers skills under .agents/skills (repository) and ~/.agents/skills (user-wide). From this repository's root, preview and install all 37 Codex skills without replacing existing paths:

python3 plugin/scripts/install-codex.py --all --dry-run
python3 plugin/scripts/install-codex.py --all

For selective and repository-scoped install, plus the Codex catalog, see codex/README.md.

Knowledge base and sources

The skills are built from a curated corpus of methods texts rather than the model's built-in knowledge. SOURCES.md is the full bibliography (150+ works). The knowledge_base/ folder holds Markdown conversions of those sources that the skills read directly when a task needs chapter-and-verse support, as fact-check does when it verifies a claim against its citation.

Contributing

Pull requests are welcome. To add a skill:

  1. Write plugin/skills/<name>/SKILL.md, following the skill authoring best practices.
  2. Add plugin/commands/<name>.md (a one-paragraph activation prompt plus $ARGUMENTS, following the existing examples).
  3. Mirror the skill to plugin/.skills/<name>.md, byte-identical.
  4. Add the Codex package at codex/<name>/ (SKILL.md and agents/openai.yaml), unless the skill is intentionally platform-specific.
  5. Add sources to SOURCES.md.
  6. Update the catalogs and badges. Install check dependencies with python3 -m pip install -r plugin/scripts/requirements-check.txt, then run bash plugin/scripts/check.sh and python3 plugin/scripts/test-codex-wrappers.py.

License

This project is licensed under Creative Commons Attribution-NonCommercial 4.0 International. The skills are intended for noncommercial scholarly and educational use.

The citation-check, literature-review, figures, tables, and figure-table-audit skills remix workflow ideas from Cheng-I Wu's Academic Research Skills for Claude Code, also licensed CC BY-NC 4.0. The instructions here are rewritten for this repository's open-science and experimental-social-science scope.

The replication-package skill adapts the structural conventions in Yusaku Horiuchi's replication-package-guide (the source for single-entry-point, compact vs. build/analyze layouts, figure/table crosswalk, paper-consistency check, correction workflow, and pre-release checklist). FAIR-principle integration and Claude Code/Codex skill packaging are added on top. Harvard Dataverse and other platform-specific upload mechanics are not included. Cite Horiuchi's guide if you publish a package built with this skill.

The spawn and research-wayfinder skills adapt concepts from Matt Pocock's skills (MIT): spawn generalizes his claude-handoff from one background successor to managed multi-session peers, and research-wayfinder reworks his wayfinder decision map from software specs to experimental design. Unmodified reference copies of the originals are vendored under third-party/mattpocock/ with his MIT license; see RECOMMENDED.md.