Agent Coach: Vibe Coding

About

Field-tested tips from everyday work with AI agents: real cases where everything looked done, yet something was missing, and what to check from then on. Agent Coach doesn’t write code for you. It shows you what to look out for when an agent does. HOW IT WORKS Every tip comes with two steps: • Read the tip: “What you do”, “Sentence for your prompt” and “How you notice it”, plus a session “From practice” with the key line marked • Quick check: one quiz question for every tip. Whatever you answer wrong lands in the “Mistakes” tab so you can repeat it 115 tips in seven chapters. CHAPTERS • AI is a tool: knowledge comes from documentation, primary sources and measurements, not from the AI’s memory • Errors from logs: real cases from daily work with AI agents, anonymized • Basic pattern: “Something is missing, but it looks complete” and why so many mistakes belong to this family • Agents in everyday work: how much you let an agent do and how to keep it safe, an agent in your own browser, presentations and your own tools for agents • How you prompt: how to scope a task, state what the result must be and ask for a cross-check • ADDP: the author’s workflow for teams of agents • Open models and your data: what changes when an agent such as OpenCode works with a local or open model, and where your data goes WHO IT IS FOR Anyone who works with AI agents day to day, for coding or other tasks at the computer, no professional background needed. The tips apply to agents such as Claude Code, Codex, Cursor or OpenCode, including open models. OFFLINE, NO ACCOUNT, NO TRACKING Everything runs on your device. No account, no ads, no analytics. Your progress stays local. No in-app purchases, no subscriptions. AUTHOR Roman Kobosil, engineer and software architect. Since 2016 he has run his own software company in Berlin, and he works with AI agents every day. Individual consulting on working with AI agents is available: get in touch via kobosil.me.