Why telling your AI coding agent "Do NOT use withOpacity" makes it use withOpacity even more (and how we fixed it across 266 Flutter tickets)
retained text
If you use Cursor, Claude Code, or Antigravity on a sizeable Flutter codebase, you have probably noticed a maddening pattern. You put a rule in your agent instructions file saying "Do NOT use withOpacity, use withValues" or "Never mock the repository under test." Two prompts later, the agent writes `withOpacity` anyway, apologizes profusely when you call it out, and does it again the next day. The culprit is how transformer attention works. Attention mechanisms do not have a hardware negation gate—when you write "Do NOT use WillPopScope," you light up the exact token neighbourhood for `WillPopScope` right before generation. Worse, when a session hits context compaction on a multi-package monorepo, procedural checklists get summarized into mush. Over the past few months, we replaced negative rules and prompt checklists with 4-part "Synthetic Scars" (Wound, Trap, Circuit Breaker, and Orthogonal Reflex), a 1976 `make` dependency graph, an isolated Adversarial Critic subagent, and a 1980 Smalltalk-style state log. Across 266 production tickets and 626 codified scars, our repeat regression rate dropped to `0.0%`. How are you structuring your agent instruction files and test gates on larger Flutter projects right now?
document fetched 2026-11-11 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xsvql5/why_telling_your_ai_coding_agent_do_not_use