AI coding assistant skill (Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and more). Turn any folder of code, SQL schemas, R scripts, shell scripts, docs, papers, images, or videos into a queryable knowledge graph. App code + database schema + infrastructure in one graph.
pip install pipxLightweight, open-source AI agent for your tools, chats, and workflows.
git clone https://github.com/HKUDS/nanobot.gitSkills Catalog for Codex
Codex uses skills to help package capabilities that teams and individuals can use to complete specific tasks in a repeatable way. This repository catalogs skills for use and distribution with Codex.
Learn more:
- [Using skills in Codex](https://developers.openai.com/codex/skills)
- [Create custom skills in Codex](https://developers.openai.com/codex/skills/create-skill)
- [Agent Skills open standard](https://agentskills.io)
## Installing a skill
Skills in [Shows active maintenance signals
2378 GitHub stars recorded
| Term | Plain-English Meaning |
|---|---|
| Harness | The workflow layer around your AI agent. It decides the next step, calls the right Skills, checks the work, and saves useful context. |
| Skill | A focused expert capability, such as tdd-guide, code-review, data analysis, writing, or research support. |
| Vibe / VCO | The canonical runtime that runs the harness. The public work entry is vibe; installed-copy updates stay on the update management path. |
| Late skill binding | The runtime binds different Skills only where they help the current work unit move forward. |
| Local skill roots | The runtime reads declared local skill roots as the only source for additional Skills. A skill must have a readable SKILL.md before it can be selected. |
| TDD / verified delivery | Work should be backed by tests, checks, artifacts, or explicit manual-review notes before completion is claimed. |
| Workspace memory | Structured project information, decisions, and evidence are stored so later sessions can continue without starting over. |
| Actual binding record | The final record of what skill was actually bound lives in work_binding, not in a discovery cache or a broad product claim. |
[!IMPORTANT]
🎯 Core Vision
VibeSkills is built for agents that need more than a tool list.
The runtime clarifies the request, plans the work, binds local Skills where they fit, records the actual binding in
work_binding, and keeps the proof needed for review or continuation.In the current release, the public entry stays narrow:
vibeis the public entry, additional Skills are discovered only from declared local skill roots, duplicate skill ids follow host root priority, andwork_bindingis the runtime record of what was actually bound.
Start with vibe. The runtime freezes the request, builds bounded work, scans the local skill roots declared by the host, binds relevant Skills when needed, verifies the result, and saves context for the next session. That matters most when the task is composite and needs more than one local Skill to finish well.
flowchart LR
accTitle: VibeSkills Harness Flow
accDescr: User intent enters the vibe harness. The harness freezes intent, builds a work model, binds helpful Skills late, verifies evidence, and preserves workspace context.
user["User Intent"]
vibe["vibe<br/>Runtime Entry"]
freeze["Freeze<br/>Requirement"]
plan["Build<br/>Work Model"]
route["Bind Skills<br/>Late"]
skills["Declared local Skills only<br/>real SKILL.md required"]
future["Local skill folders can grow<br/>while the public entry stays the same"]
verify["Verify<br/>Tests + Evidence"]
memory["Remember<br/>Workspace Context"]
user --> vibe --> freeze --> plan --> route --> verify --> memory
route --> skills --> verify
route --> future --> verify
classDef core fill:#ede9fe,stroke:#7B61FF,stroke-width:2px,color:#1f1147
classDef stage fill:#e0f2fe,stroke:#0284c7,stroke-width:1.5px,color:#0c4a6e
classDef proof fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef user fill:#fff7ed,stroke:#f97316,stroke-width:1.5px,color:#7c2d12
class vibe core
class freeze,plan,route,skills,future stage
class verify,memory proof
class user user
| Signal | What it means |
|---|---|
one entry | Start with vibe; use update to refresh the same installed skills directory. |
late skill binding | Skills are attached after the work shape is clear, not used as the control plane. |
local skill roots | The runtime checks declared local skill roots and only considers entries with a readable SKILL.md. Duplicate skill ids keep the highest-priority root active. |
actual binding record | The runtime truth for selected skill provenance lives in work_binding, even when discovery or benchmark artifacts are also written. |
proof trail | Tests, checks, artifacts, or manual-review state support delivery claims. |
memory plane | Requirements, plans, decisions, and evidence survive the chat window. |
The normal closeout path should stay small: prove the governed runtime, entry truth, execution proof, release consistency, and repo cleanliness before reaching for wider audit gates.
vibe| Image Workbench | Build a GPT-image workspace for prompt chat, reference uploads, and real image generation. | Turns the idea into a product scope, UI/API tasks, workflow checks, and screenshot review. |
| Video Editing Pipeline | Recut a rocket moon-landing history clip into a short-video style edit. | Breaks the media work into caption, music, pacing, render, and review passes, with rough edges recorded plainly. |
| ML Experiment + Paper | Build a face-recognition ML demo and turn the run into a paper. | Guides dataset and model choice, training, evaluation, figure generation, and LaTeX compilation. |
A useful demo should show both the result and the path that produced it:
flowchart LR
accTitle: VibeSkills Practice Demo Flow
accDescr: A user goal moves through scope confirmation, planning, skill handoff, execution checks, and an openable result.
goal["Plain Request"]
freeze["Confirmed Scope"]
plan["Work Plan"]
route["Skill Handoff"]
work["Build / Analyze / Render"]
proof["Checks + Evidence"]
show["Openable Result"]
goal --> freeze --> plan --> route --> work --> proof --> show
classDef goal fill:#fff7ed,stroke:#f97316,stroke-width:1.5px,color:#7c2d12
classDef flow fill:#e0f2fe,stroke:#0284c7,stroke-width:1.5px,color:#0c4a6e
classDef proof fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
class goal goal
class freeze,plan,route,work flow
class proof,show proof
Inspired by the VibeSkills 3.1.0 community practice cases: a GPT-image workbench, a video-editing run, and an ML experiment that produced a paper. The best examples link to concrete outputs: a running app, a rendered clip, a compiled paper, or the commands and evidence used to produce them.