Jan Mikoláš
Agent work
A bounded implementation loop
- Context, rules, and a concrete task
- Research and proposed approach
- Implementation of a small change
- Running the prescribed checks
Back to skills
Way of working
AI-assisted development
AI speeds up research, design, and well-bounded implementation for me. Responsibility for the brief, architecture, security, and result always remains mine.
A good result begins with good context.
Before involving an agent, I prepare the goal, scope of change, repository rules, relevant architecture, constraints, and the method of verification. The agent then works with a concrete problem in a clearly bounded environment, not merely a general prompt.
I split work into small, verifiable steps: research, design, implementation, and review. Custom instructions, skills, and MCP integrations help repeat established workflows; they do not replace technical judgement or review.
A controlled process
From the brief to a verified result, not blind generation.
-
1. Context
An agent needs to know the environment and boundaries
The brief includes not just the requested change but repository rules, conventions, existing solutions, risks, and verification commands. Without this, even a good model is left mostly to guesswork. -
2. Work splitting
Small tasks with an explicit outcome
Research, design, code changes, and verification do not have to be one step. Splitting work makes it possible to catch a wrong assumption earlier, compare alternatives, and keep a change within a sensible scope. -
3. Reviewing a proposal
A typical example: unnecessary abstraction
An agent may propose another service, repository, or layer simply because it recognises a familiar pattern. If the new abstraction does not separate a real responsibility, I simplify the proposal or have it reworked to follow the project architecture. -
4. Verification
Code goes through the same checks as a manual change
I compare the result with the brief and check it through relevant tests, PHPStan, coding standards, architectural rules, and when needed, manual or browser verification. -
5. Responsibility
What I do not delegate to AI
I do not let AI independently decide business rules, security, data changes, production actions, or acceptance of a result into a project. These choices require human context and accountability.
Human–agent collaboration
The agent accelerates work; I keep direction and quality.
I use AI as a tool for focused work on a clear brief. The more precise the input and verification, the more useful the result — and the less time is lost to corrections.
Sensitive information, access rights, destructive operations, and production decisions remain under direct human control.
My responsibility
Decisions and final review
- Priority and business meaning
- Architecture and security
- Reviewing the diff and side effects
- Approval for merge and deployment
A public example of the workflow
Repository instructions and automated verification
AGENTS.md as context for an agent
The public project includes explicit repository instructions. It is an example of providing an agent with boundaries, rules, and a verification process instead of a vague brief.
Automated checks as a safeguard
AI assistance feeds into the same tests and quality checks as manual development. The repository workflow gives the result independent feedback.
Tools and practices
AI as part of the development workflow
- Codex
- ChatGPT
- Claude
- Custom skills
- MCP
- Agent workflow design
- PHPUnit
- PHPStan
- CI/CD
Looking for development that uses AI without losing control?
I use AI for faster, more focused work, while keeping every change within architecture, quality checks, and accountability for the result.