Asking coding agents to check their work allows for an easy way to add self-reflection and works correctly across models like Codex or Claude Code but, what would to be counterintuitive, really is just looking at the last output for inaccuracies. One would think that both are set up to critique themselves. Instead, the agent actually finds errors or gives feedback that it couldn’t find anything. Even adding “check your work” to the default prompts that are fed into the agents, such as using AGENTS.md and asking on the next query, this technique can find inaccuracies. I would say asking “check your work” is the easiest way to make the agent do self-reflection without explicitly adding it.