Hand your coding agent the engineering work that repeats, and spend your energy on what only you can decide.
Coding agents can write code. The rest of the job still falls on you: tracing the bug, planning the change, picking the model, reviewing the diff, proving it works and writing it all up.
Orbyn is a free library of 39 connected skills for Claude Code, Codex and Cursor that hands that work to the agent. You say what you want in plain words. The agent picks the method, splits the work across the right models, checks the result and stops for your approval before anything leaves your machine.
Our goal is simple: automate everything that can be automated, so engineers can put their energy somewhere better.
git clone https://github.com/orbyn-dev/orbyn-fullstack-skills.gitOpen that folder in your agent and paste this, with your repository's URL:
Set up the nested workspace for
<target repository URL>. Read our engineering policy and the target's own instructions and skills. Preserve existing work, clone or safely reconcile the target underrepo/, check the workspace, and complete the setup yourself. Do not install or migrate shared skills into it.
It works with Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor, Windsurf, Cline, Roo, Kiro, Zed, OpenCode, Antigravity, Amp, Junie and Pi; see supported agent hosts.
Then ask for ordinary work:
In repo/widget, reproduce this failure and fix it.
Threat-model the new upload feature before we build it.
Draw our stack as a diagram and show how a request moves through it.
You don't choose skills or flags. It needs Git and Python 3.11+, installs nothing into your project, and needs no Orbyn account.
| You used to | Now the agent |
|---|---|
| Work out which method fits the task | Routes the request to the skill that owns it |
| Read unfamiliar code to find a cause | Traces the source and reproduces the failure |
| Pick models and limits for each job | Reads this repository's settings before it starts |
| Use your strongest model for everything | Leads with Opus or Sol, codes with Sonnet or Terra, and gives mechanical steps to Haiku or Luna |
| Trust a summary of what changed | Reads the real diff and reports what ran and what did not |
| Remember the security review | Threat-models a design, audits source, fixes and verifies separately |
| Write up how a PR was built | Fills in tools, models, sources and evidence |
| Track who to credit | Pins each source to a commit and generates CREDITS.md |
You keep the decisions. Commits, PRs, merges and deployments need your explicit yes. No file in a repository can grant an agent permission.
Every request follows one path, and checks hold it to the rules at each step.
Blue is the path of one request. Grey shows what guides it, and dashed lines show checks and release.
Each task resolves its repository's settings first, then moves through planning, delegation, review and verification. You're asked to review it only once verification and the PR gate have passed.
The agent loads only what the job needs, so 39 skills don't crowd its context. Small edits skip the ceremony. Rules run as code wherever they can, so CI catches drift that written policy would miss.
- Plan and design: product discovery, architecture, data migrations, stack choices and statecharts.
- Build: APIs, web and mobile UI, AI features and 13 languages.
- Prove: tests, CI design, adversarial review and threat models.
- Investigate: source tracing, runtime forensics and shipped binaries.
- Show your work: architecture diagrams, faux product demos, code-only films and lyric videos, and social posts with previews.
See every skill, or start from the task map.
Every PR runs the library checks and tests on Python 3.11 and 3.13. We also publish small evaluations, including the results that were weak.
Skills guide an agent; they can't guarantee correct software. Our eval cases are mostly static, and we haven't measured overall quality gains across models. The guide states the limits of each result.
Orbyn works with your stack: it reads your schema, rules and owners from your source, and your project's own instructions always win. The library is MIT licensed, and we built it from other people's open work, all named in CREDITS.md.
We expect skills to be built with agents, and a PR states how: the tools, models and sources used. Read CONTRIBUTING.md to add yours, and report vulnerabilities through SECURITY.md.
If this saves you time, star the repository and show someone else what your agent did. Maintained by Anh Dang for Orbyn Engineering in the orbyn-dev organization.