Prompt Patterns for AI Coding Agents: Plan-First, Spec-First, and Anti-Patterns

The agent one-shots your feature on Monday and quietly rewrites your auth module on Tuesday. Here is the workflow that keeps the speed without the wreckage: plan-first prompting, reviewable diffs, and verification the agent runs itself. With interactive tools to grade your own prompts.

You correct the agent, it apologizes, and ten minutes later it makes the same mistake. That is not a model problem; it is a memory problem. Here is how the context window works as a budget, what belongs in CLAUDE.md and AGENTS.md, and when to restart a session instead of patching it.

An agent that has to prove its work catches its own mistakes. One that only has to look right ships them to you. This is how to build the feedback loop: the verification ladder, test-first prompting, and letting the agent iterate to green on its own without weakening your tests.