No new app. No new habit. Just better English while I work.

My coding agent reviews every prompt. An English check comes first, then the original task. I started from a single line: “Always review my English before answering.” The reviews were unstable. One run caught a missing article, the next handed me a long read about writing. Then the agent kept forgetting. Despite “always” in the instruction, reviews still went missing. I wanted it to run without me. Silence had a green light. The instruction allowed it when there was nothing of mine to check, like a message in another language or pasted content. So a forgotten review looked exactly like a correct one. Forbidding silence was the fix. Every reply starts with exactly one status, no exceptions: × English errors found ✓ No English errors – English review skipped Long mixed messages were the trickiest. The agent kept skipping the review when my own English was there. It took three fixes: • Checking is the default. Skipping has to be earned. • Remove everything that isn't mine first: pasted text or code, other languages. • Add worked examples like “Compare these” plus two pasted texts. A strict format kept it stable: • Shortest fragment. • Correction. • One-line reason. About 1,100 tokens per session. That's the cost of the instruction, and I've noticed no quality drop. Even in thinking mode, only the first thoughts go to the review. Did my English actually improve? Yes, especially articles. My guess is frequency, since they show up in almost every sentence. This is my only global user-level instruction, in CLAUDE.md/AGENTS.md. The full instruction → https://gist.github.com/resivalex/bdfd0769672b9c8a954a49f5fb9c2b06