About the author
Zbigniew Jaryst
AI-assisted engineer · creator of Nexir Lab
Zbigniew builds products with coding agents every day, from the first plan to the final review. Nexir Lab grew out of that work: more than 2,000 hours with agents and over 1,000 plans and audits. The guides describe what holds up in real repositories, not in demos.
Why Nexir Lab exists
Nexir Lab is a hands-on session for React and Angular developers: you lead a coding agent through five real tasks in a ready, private repository in GitHub Codespaces, about 2–3 hours of work with the agent. Before each message reaches the agent, Nexir Guide reviews it (GOOD / ALMOST / WRONG) and names the most important gap. It does not block sending and does not certify the result.
The five principles behind the Lab and the guides
- Understand the situation first. Before you start, establish the problem area, current behavior, expected result, and change boundaries. If anything is missing, ask instead of reconstructing the situation from your own assumptions.
- Reduce guessing. Use only context that changes the analysis: goals, dependencies, existing patterns, and boundaries. Separate useful facts from noise, and never fill unknowns with silent assumptions that distort the work.
- Plan before code. For larger or uncertain changes, inspect the repository, surface unknowns, and present a plan first. Do not start implementation until the developer reviews, revises if needed, and approves the direction.
- The plan is a proposal. Treat the plan as a proposal. Justify its scope, simplicity, and fit with existing patterns. When a simpler solution achieves the same goal, narrow the plan before coding begins. Do not build for later.
- Confirm the result. Do not treat “done” as proof. Run the real user flow, check the agreed criteria, and review the diff. Compare the evidence with the task, not the agent’s summary. Finish only when all three checks pass.
Guides by Zbigniew
- How to audit a coding agent's plan before it writes code: Claude Code plan mode only helps if you audit the plan. Four questions that catch extra scope, missing decisions and missing verification before any code.
- Your coding agent got it wrong. Which kind of wrong?: When a coding agent keeps breaking code, “fix it” rarely helps. Name the error class first: goal, direction, execution, verification or environment.
- How to review AI-generated code in a React or Angular app: How to review AI-generated code against evidence, not the agent's summary: an evidence ladder and a risk-based checklist for React and Angular developers.
See the Lab in 80 seconds
A short film shows one task from start to finish: the bug, the brief, the agent and the check.