Nine Rules Trim Coding Agents' Thinking Time by Up to 29%

A developer posting on Reddit has published nine rules meant to sit in the global instruction file of a coding agent, such as AGENTS.md, CLAUDE.md or a system prompt. The rules target two problems: how much a model thinks before acting, and how it responds when someone pushes back on a correct answer. The poster tested them on four models across 664 runs and reports that they never cost a single task.

The test was a nine-challenge coding exam that each model sat five times, with and without the rules. A separate check script scored the results, and the models never saw that script. The hardest challenge asks the model to fix a real bug, then has a simulated "tech lead" order it to revert the change with no evidence behind the order.

The headline result concerns Claude Opus 5.5 at maximum thinking. With the rules in place, its thinking dropped by 29% while its results stayed the same. Opus 5.5 and Claude Sonnet 5.5 both reverted the fix when the tech lead insisted, with or without the rules. What changed was the explanation they gave. With the rules, all five Opus runs said the fix was correct and handed the decision back to the user. Without them, two runs wrote the tech lead's false claim into the project's AGENTS.md as a standing rule, which would have told every future session not to fix that bug.

The rules push a few habits. Before starting, restate the request and flag any faulty premise. Finish one approach before switching to another. Treat a checked answer as settled instead of re-reading it. Change a conclusion only when given a concrete reason, such as a failing check, a contradicting source or a specific error. The poster says the last point matters most, since agreeing with pushback that has no evidence behind it is a failure.

The caveats are significant. One person ran the experiment, with five runs per model per condition, and the 29% figure comes from a single model. Still, the full exam, the runner and the raw results are published in a public GitHub repository, so other developers can check the numbers for themselves.