Stop maxing reasoning effort: iterate on low, audit on max
johnlindquist · x · 2026-09-17
John Lindquist shares a practical strategy for using reasoning models: instead of cranking reasoning effort to max, iterate fast with models on "low effort" once you deeply understand your product, audience, and goals — then audit on max effort.
The linked MEGA article explains what reasoning effort actually does: it controls how many tokens go into the model's internal reasoning trace, not a hard cap. At higher effort, models spend more time on extra requirements, dig harder for the code that needs changing, and catch edge cases you didn't think of. With vendors now shipping model families each with their own effort dials (GPT-5.6 Terra alone has low through max), strategically allocating effort beats blindly maxing it.
More from coding & agent
- Rene launches: an iMessage agent that chases invoices and collects cash for you — EXM7777 · 2026-09-17
- Solo dev builds STARBATTLE in 7 weeks with all code, art and audio AI-generated in native C++ — AIandDesign · 2026-09-17
- Dev demo: grouping tabs, files and terminals by project with swipe switching — nateparrott · 2026-09-17
- Google AI Studio Tips: Using generate_image to Make Apps Stop Feeling Vibecoded — m4rkmc · 2026-09-17
- Tencent open-sources TeamAI-CLI after 6 months of internal use to turn team knowledge into one git repo for agents — Roger_M_Taylor · 2026-09-17
- RoMem uses continuous phase rotation to expire stale agent memories, 2-3x better temporal reasoning — Roger_M_Taylor · 2026-09-17