I'm roebi — Robert Halter, Swiss software engineer with years of experience in test automation and CI/CD.
A while back I started asking: what if the way we work with AI could be as structured and reusable as good code?
That question led to New Work with Skills — a methodology where every task you give an AI agent is defined as a versioned, portable SKILL.md file. Not a vague prompt. A skill.
This is not just theory. It is the way I actually work.
Based on the skills I created and used in my own framework, sw-dev-agent-framework, I refined a practical process for building software in a disciplined, repeatable way. By using and improving those skills in real projects, I developed a workflow that is both specification-driven and TDD-driven:
- Start from a clear requirement and a concrete specification
- Group requirements into named, dependency-aware groups
- Order groups by dependencies: independent groups first, dependent groups only when their prerequisites are satisfied
- Turn the approved requirements into architecture and implementation decisions
- Drive development through failing tests first
- Validate each step before moving on
- Review, refine, and release with traceable intent
So the model is not: "ask AI to code" in a loose way. It is: define the work as a skill, execute it consistently, and improve the skill over time as the process matures.
You don't prompt. You invoke a skill.
Skills are composable. Skills are auditable. Skills ship.
1. Define — Write a SKILL.md that captures what the agent should do, how, and when it's done.
2. Invoke — Any LLM agent reads the skill and executes it deterministically.
3. Compose — Skills call other skills. Complex workflows from simple building blocks.
4. Version — Pin skills like dependencies. Reproducible today, next year, on any model.
5. Improve — Use them in real projects, learn from the outcomes, and evolve the workflow.
This is the agentskills.io open specification — and it matches the process I developed while building and refining sw-dev-agent-framework: specification-first, TDD-first, and continuously improved through repeated use.
The requirements phase in that framework is especially important:
- requirement groups are split into clear, named clusters
- each group declares its dependencies explicitly
- approval follows dependency order, not arbitrary ordering
- later design and TDD phases can build from an already structured plan instead of guessing what's first
In the order shown on my GitHub profile:
- sw-dev-agent-framework
- agent-skills
- ai-consolidation-timeline
- new-work-skills
- agent-security-skills
- aider-skills
| 🌐 new-work-skills.dev | The methodology, skill cards, live examples |
| 📝 Blog | What I'm learning, building, and discovering |
| 🤖 agent-skills | My personal skills library — open source |
| 🛠 aider-skills | Agent Skills integration for aider |
| 🧰 sw-dev-agent-framework | Specification-driven, TDD-driven software development with agent skills |
roebi · Robert Halter · New Work with Skills · new-work-skills.dev · Switzerland



