On 27 August 2026, I gave a talk at MAINDSET in Žilina called Ako si nastaviť workflow pri programovaní s AI.
The talk focused on the work around AI-generated code. Coding agents can implement a well-defined change quickly, but that speed is useful only when the repository, task, verification process, and human feedback loop are prepared for it.
Open the public presentation. It includes English and Slovak versions, presenter notes, and additional resources in the side panels.
What the presentation covered
The presentation starts with technology selection. A project should begin with product constraints and research, not with loyalty to one framework. It then compares two repository shapes: a small Deno and Fresh website, and an anonymized Nx and pnpm monorepo pattern for a larger system.
From there, the focus moves to repository context. AGENTS.md holds durable rules, skills package recurring workflows, and source documents carry the architectural and product detail. Larger changes are divided into a specification, an implementation plan, and a task list with explicit validation commands. I also covered OpenAI's ExecPlan approach as a single-file alternative.
The final part deals with verification and steering. Automated checks, browser inspection, and human review solve different problems. A local tunnel provides a fast preview while work is still changing, while a pull-request preview gives a stable deployed checkpoint. This also makes it practical to monitor agents and review visual results from a phone without pretending that the phone is a good place to write code.
Resources behind the talk
The presentation builds on several articles I had already written about AI-assisted delivery:
- Fast Code, Slow Delivery
- AI Coding Did Not Fix Slow Delivery. It Exposed It.
- AI-Native Delivery Is a Team Sport
- From formula to production in a few hours
- The Interviewer Problem
The external references include OpenAI's guidance for AGENTS.md, skills, and ExecPlans, together with documentation for Fresh, Nx, Playwright visual comparisons, and Cloudflare Tunnel. I also referenced the Amplifying study on agent tool preferences and used Taste Skill while designing the deck.
The practical takeaway
The agent can write the implementation. The more durable advantage comes from designing a development system that keeps its work understandable, verifiable, and easy to steer.