Codex Skills Library

Codex Skills Library collects 19 installable workflows distilled from repeated research, software, and technical operations work. A skill is closer to a lightweight SOP than a prompt: it tells Codex what to inspect, which checkpoints matter, what artifact to leave behind, and when to stop for a human decision.

Use one skill for a focused job or compose several across a longer project. Each installable directory keeps its instructions, declared requirements, license, and provenance together.

Research and paper work make up the largest part of the current catalog, but this is not a research-only library:

• Research and papers · 10 skills: experiments, evidence, recent-paper reading, writing, review, rebuttal, figures, and presentations.
• Apps and UI · 5 skills: product design, feature delivery, debugging, and application releases.
• General methods · 2 skills: clarify an underspecified plan and search before building.
• GitHub and Codex operations · 2 skills: publish a clean repository or restore missing Codex Desktop sessions.

Independent community project; not affiliated with or endorsed by OpenAI.

Featured · Feishu paper reading

Turn a research question into an evidence-grounded, read-back-verified Feishu brief, not a link dump.

feishu-paper-reading owns the whole path: broad candidate retrieval, primary-source verification, full-text and appendix reading, an evidence ledger, cross-paper synthesis, report validation, checkpointed Feishu publication with no blind duplicate retries, and read-back verification. When no writable route exists, it can guide a least-privilege official setup without bundling credentials or silently taking over an account.

Explore the workflow, real output excerpts, setup behavior, and install prompts →

Start with one skill

Codex includes the $skill-installer system skill. Give it the GitHub URL of an individual skill directory. Choose the example closest to the work in front of you.

Research or coursework

$skill-installer Install https://github.com/sidiangongyuan/codex-skills-library/tree/main/skills/experiment-planner

On the next turn, try a concrete request:

$experiment-planner Turn my image-classification course project into a one-day pilot with baselines, metrics, resource assumptions, and a stop/go decision.

A useful result should contain a falsifiable claim, comparisons, the smallest pilot, expected signals, and a decision gate, not just a longer brainstorm. The worked example shows the shape of that artifact without pretending that unrun experiments produced results.

Application debugging

$skill-installer Install https://github.com/sidiangongyuan/codex-skills-library/tree/main/skills/app-bug-forensics

On the next turn:

$app-bug-forensics Trace this intermittent timeout from the visible symptom through logs and the request path. Identify the root cause before changing code.

A useful result should distinguish observations from hypotheses, identify the failing boundary, and leave behind a focused regression check.

There is no need to install all 19 skills. Replace experiment-planner with any name in the catalog. The installed skill is available on the next turn. Review its SKILL.md, LICENSE, requirements, and provenance before using it in a sensitive environment. See the single-skill installation details for private forks and command-line alternatives.

The repository installer supports selected or bulk installation and has no runtime dependencies outside the Python standard library. It defaults to the shared user-level directory $HOME/.agents/skills.

git clone https://github.com/sidiangongyuan/codex-skills-library.git
cd codex-skills-library
python scripts/install.py
python scripts/install.py --all --dry-run
python scripts/install.py --all

Running the installer without selection arguments only prints the catalog and usage. --all is always explicit. To install a smaller set:

python scripts/install.py \
  --skill experiment-planner \
  --skill app-bug-forensics \
  --skill search-first \
  --dry-run

python scripts/install.py \
  --skill experiment-planner \
  --skill app-bug-forensics \
  --skill search-first

Existing skill directories are skipped by default. Use --replace only after reviewing the dry run. Use --target <directory> to install elsewhere. The installation guide covers Windows paths, manual copying, legacy --codex-home compatibility, and troubleshooting.

Start from the work in front of you

• A vague plan or an unfamiliar task: grill-me → search-first clarifies the decisions that matter and checks maintained tools or prior work before anything custom is built.
• A software coursework or capstone project: grill-me → search-first → app-feature-craft turns a vague brief into a bounded project, checks existing tools, and carries the chosen path through implementation and verification.
• A deep-learning research idea: grill-me → search-first / research-evidence → experiment-planner separates the claim from the hunch, checks prior work, and ends with a pilot-first experiment matrix.
• A recent-paper reading digest: feishu-paper-reading searches a broad candidate pool, deeply reads the selected papers, preserves short original-language evidence, synthesizes the set, and publishes a verified Feishu report. If no verified connection exists, it guides the official least-privilege Feishu setup step by step before publishing; Markdown fallback is reserved for a declined, incompatible, or unsuccessful setup.
• A paper before submission: paper-section-playbook → paper-refinement-skills → paper-review-panel structures the argument, tightens it without inflating claims, and exposes decision-changing review risks.
• Official reviews and rebuttal: rebuttal-response-skills takes over only after reviews arrive, mapping exact concerns to evidence and reviewer-specific responses.
• Figures, tables, and a research talk: paper-framework-figure-studio-pro → paper-visual-craft → paper-share-html plans the visual story, refines the evidence, and prepares an audience-ready presentation.
• An application or open-source release: app-feature-craft → app-bug-forensics → app-release-readiness builds the feature, traces failures to root cause, and verifies what is actually shipped.
• Repository publishing or Codex recovery: github-project-release audits a project before publishing, while codex-session-restore handles a specific Codex Desktop session-recovery problem after provider switches.

Browse by goal

The concise table below is generated from skills.json. Requirements, example prompts, licenses, and pinned source revisions remain in the generated full skill catalog.

For example prompts, dependency notes, licenses, and pinned upstream revisions, open the generated full skill catalog. Attribution and source treatment are recorded in NOTICE.md.

Use a skill

Explicit invocation is the clearest way to make a workflow reproducible. Start with a general method when the task itself is still unclear:

$search-first Find maintained tools and libraries for this dataset utility before we decide to build it ourselves.

For an application failure:

$app-bug-forensics Diagnose this intermittent provider timeout from the UI state through the request path. Report the root cause before changing code.

For research planning:

$experiment-planner Turn this idea into a pilot-first experiment matrix with a falsifiable claim, baselines, diagnostics, and a stop/go gate.

Before submission, use the review panel to expose decision-changing paper risks:

$paper-review-panel Review this draft before submission. Separate decision-changing evidence gaps from issues that can be fixed with writing.

Most included skills allow implicit invocation when their description closely matches the request. Skills whose default workflow publishes externally may require explicit $skill-name invocation; their agents/openai.yaml records that policy. Explicit invocation is preferable for repeatable handoffs.

Skills can be composed in sequence, but each stage keeps a clear owner. See workflow recipes for complete examples and handoff points.

Contribute

Contributions are welcome from any domain. A proposal issue is optional; a pull request may add a useful skill directly. New skills must include:

• a focused SKILL.md with a matching lowercase kebab-case name;
• an example prompt and complete requirements in skills.json;
• an applicable license inside the installable skill directory;
• original-work or pinned-upstream provenance; and
• safe defaults for destructive, publishing, or externally visible actions.

Do not include credentials, private conversations, unpublished project material, machine-specific paths, datasets, checkpoints, or redistribution rights that are unclear. Read CONTRIBUTING.md for the template, validation commands, and review criteria.

License and safety

Original project material is licensed under the MIT License. Third-party material remains under its upstream license and copyright; every installable skill carries the license text that applies to it. See NOTICE.md for provenance labels and fixed source revisions.

Skills can influence tool use and may include executable helpers. The library installer only copies files and never executes installed helpers, but users should still inspect skills before granting them access to sensitive data, credentials, publishing surfaces, or destructive tools. Report security issues through the process in SECURITY.md.