Manage Reasoning, Don't Beg for Chain of Thought
The output is under control, but many still beg the model to "think step by step". Modern reasoning models reason internally on their own. The application needs the right result, evidence, and observable actions, not a long monologue about how the model thought.
A useful instruction depends on the task, whereas the redundant one is almost always the same.
- A simple lookup - "answer immediately, use account_read". Redundant: "reason as deeply as possible".
- Ambiguous diagnosis - "first gather the missing facts, then pick the route". Redundant: demanding that the internal reasoning be published.
- A long agent task - "keep a plan, check tool results, continue until done or blocked". Redundant: unlimited autonomy with no stop condition.
- A high cost of error - "check the result against explicit criteria and request approval". Redundant: trusting the model's phrase "I checked everything".
The companies don't fully agree here, and that is worth accounting for.
- OpenAI. For GPT-5.6 it advises precise instructions and notes: the model recovers intent better, so prescribing every step is often unnecessary. Domain context, hard boundaries, approval, and success criteria are still mandatory.
- Anthropic. The advice depends on the thinking mode and generation. For Opus 5 a redundant self-check instruction can trigger over-verification.
- Google. Gemini 3 prefers direct instructions and may over-analyze cumbersome old techniques. Depth is tuned to the task.
The common denominator is to ask not for confidence but for a verifiable signal.
Ask for an observable check. "State the source_id you used", "run the tests", "compare the result against the schema", "don't write before approval" give a verifiable signal. "Be confident" doesn't.