A practitioner shares seven practical rules for getting reliable results from AI coding agents, covering context, documentation, testing, UX review, and model collaboration.
Adapted from @PawelHuryn7 truths about working with coding agents, learned the hard way: 1. /goal is not enough. Explain the why + strategic context (market segments, value proposition, how you make money, etc.). Progressive disclosure. That's what enables agents to make better, more autonomous decisions. 2. To review code, agents need something to review it against. In particular, it's critical that agents document the intent, for example, roles and permissions. Without it, they may not be able to distinguish a bug from a feature. 3. Documenting the intent is also a prerequisite for reliable tests. 4. Even with browser automation, agents still can't review UX effectively, especially dynamic elements, micro-interactions, window resizing, and so on. Asking the agent to take screenshots and inspect them (exploratory testing) helps a lot, but in 2026, you still need a human eye. 5. Human code review is necessary only in the absence of strategic context and documentation. 6. If you must review something technical, review the architecture and procedures - not the syntax. How can you make your procedures, e.g., release, more deterministic? 7. You will get way better results by letting Astra and Opus/Fable work together. Neither OpenAI nor Anthropic would tell you so.
