Agent Skill Development

When Your AI Coding Agent Writes Code – But Not YOUR Code

AI coding agents like Claude Code, GitHub Copilot or Cursor are impressively productive. But out of the box, they know nothing about your architecture, your frameworks, your conventions or your domain. The result: generic code, repeated corrections, and developers who spend more time reviewing than they save.

Another Approach: Teach The Agent How You Work

Agent Skills package your team’s know-how into reusable, versioned instructions, scripts and references that the agent loads exactly when it needs them. Instead of explaining your conventions in every prompt, the agent knows them – consistently, for every developer on the team.

I build Skills like these myself. The AI Unified Process plugins for requirements, use cases, entity models, Vaadin, jOOQ and testing are used in real projects, and everything I learned building them goes into the Skills I develop for you.

Why Prompting Alone Is Not Enough

Knowledge Stays In Heads
Every developer writes their own prompts. Good practices are not shared, and the agent’s output depends on who is asking.

Context Gets Lost
Long instruction files are ignored or only partially followed. The agent needs the right knowledge at the right moment – not everything all the time.

Quality Is Not Measured
Without tests, nobody knows whether a prompt or rule actually improves the result – or whether it stopped working after the last model update.

Skill Development: From Know-how to Reliable Agent Behavior

Analysis – Identify the tasks where your agents fail or need the most corrections
Knowledge Capture – conventions, architecture rules and patterns from code, documentation and your experts
Skill Design – Structure the knowledge into focused Skills with clear triggers, references and scripts
Evaluation – Write evals that measure whether a Skill improves the result, compared against a baseline without it
Rollout – Distribute the Skills to your teams as plugins via a marketplace or registry
Maintenance – Keep Skills up to date when models, frameworks or your conventions change

Your benefit

Consistent
Every developer gets the same quality, regardless of prompting skills.

Faster
Fewer correction loops, less review effort, more code that fits on the first attempt.

Measurable
Evals show what a Skill delivers instead of relying on gut feeling.

Sustainable
Your know-how is versioned, documented and outlasts team changes.

The Principles Behind It

Focused – One Skill, one job. Small Skills trigger reliably and stay maintainable
Evaluated – No Skill without evals that prove its value
Progressive Disclosure – The agent loads details only when needed, keeping the context lean
Tool Independent – Based on the open Agent Skills format, usable across agents
Versioned – Skills are code: reviewed, tested and released like any other artifact
Built from Practice – Derived from your real code and real problems, not from generic templates

When This Approach Makes Sense

Your teams use AI coding agents, but the results do not match your standards
You use in-house frameworks, libraries or architecture rules the agent cannot know
Developers repeatedly correct the same mistakes
You already have instruction files or Skills, but you don’t know whether they work
You want to establish AI-assisted development across several teams

How I Support You

Skill Audit
I review your existing Skills, rules and instruction files: Do they trigger reliably? Do they improve the result? You get a concrete list of improvements.
Workshop
In a one-day workshop, your team learns how to design, write and evaluate Skills – using examples from your own codebase.
Custom Skill Development
I develop Skills tailored to your stack, frameworks and conventions, including evals and documentation, ready to roll out to your teams.
Maintenance
Models and frameworks evolve. On request, I keep your Skills up to date and re-run the evals after every relevant change.