Reasoning and sycophancy: think hard, frame neutrally, force a number

Concept 5: reasoning, or "think hard"

Until about 2023 the standard advice for hard prompts was "think step by step". That advice is now mostly obsolete: modern models have built-in reasoning modes you can invoke directly. ¹

How to invoke it: ask in plain language, "think hard" or "think carefully before answering", the portable move that works in every modern chat tool; use the thinking-mode toggle where one is offered; and on some products the tool decides on its own when a question is hard enough. When extended thinking is on, the model can think for many seconds, on hard problems sometimes more than ten minutes. It is not typing slower; it is internally exploring multiple approaches, checking its own work, and only then writing the answer you see. ¹

A 2025 METR study tracked the longest task a frontier model could reliably complete. In mid-2024 a leading model handled tasks that take humans around seven minutes. By early 2025 that was up to roughly an hour, and the length it measures has been doubling roughly every seven months. The implication: hand AI real, hard tasks, not just easy ones. ^{1 2}

A power-user pattern that uses this well: ¹

I'm choosing between two cars. Attached: spec sheets for both,
my insurance quote for each, and a spreadsheet of my driving
patterns over the last six months.

Read everything. Think hard. Then tell me:
1. The three trade-offs that actually matter for my driving pattern.
2. Which car you'd choose and why.
3. Under what conditions your recommendation flips.

Three things this prompt does: it loads the relevant context, it explicitly invokes thinking, and it asks for structured output instead of a wall of prose. All three are habits. ¹

When not to use thinking mode: quick lookups, summaries of a paragraph, casual brainstorming. It is slower and uses more of your usage budget. Save it for the questions where you would have wanted a human to take their time. The trade is real: a few minutes of compute for something you would have spent half a day producing. And re-test your assumptions about what AI can do every six months. They will be wrong. ¹

Concept 6: sycophancy and how to neutralize it

AI models are trained on human feedback, specifically on which responses got a thumbs up. Across millions of users, agreeing with people gets more thumbs up than disagreeing. The result: models are biased toward telling you what you want to hear. A November 2025 Washington Post analysis of 47,000 ChatGPT conversations found the model opened with an affirmation ("yes", "correct" and similar) about 10 times more often than it opened with "no" or "wrong". ^{1 3}

You can verify this yourself with opposite framings: "Don't you think remote work is better than office work?" gets agreement and reasons; "Is it true that office work is more productive?" gets agreement and reasons. ¹

The fix is not magic. It is neutral framing, at two levels, surface ("don't you think X?") and subtle ("find evidence that X works"): ¹

The pattern: any phrasing with a verb like find, defend, confirm, prove, support hands the AI a conclusion before the question. Replace with evaluate, compare, critique, find any, list both sides. The general rule: lay out two options without hinting at preference, then ask for pros and cons of each. If you find yourself writing "isn't X true", rewrite it as "to what extent, if at all, is X true?" ¹

The objective-rubric pattern

A founder asked: "I have a great business idea, mobile tie-dyeing for kids' birthday parties, critique it." The AI praised it warmly. The founder tried again: "Analyze this idea objectively. For each of the following, score 1 to 10 and justify: (1) is there a real problem here, (2) is there a market willing to pay, (3) is there a competitive advantage, (4) what's the unit economics, (5) what are the top three reasons this fails." The same AI gave the idea 8 out of 100 and explained, in concrete terms, why. Same model, same idea, opposite verdicts. ¹

A rubric is a list of specific things to check, each scored or answered separately. Without one, ambiguous criteria collapse into "great work". With one, specific criteria force the AI to actually look. ¹

Force a number

For each criterion, require a score on a fixed scale, 1 to 5 or 1 to 10, with a one-sentence justification. This works for two reasons. First, what the number does to the AI: a model that wants to please you can call your draft "strong" without committing to anything; asked to pick between 6 and 7, it has to commit, and committing forces it to look. Scores tend to come in lower than the prose suggests, because the prose was sycophantic and the number is not. Second, what the number does for you: "strong" and "could be tighter" cannot be compared, prioritized or tracked; a 4 and a 7 tell you which criterion to fix first, and today's 6 versus last week's 5 tells you whether the second draft improved. ¹

Grade each criterion out of 10, with a one-sentence justification. Then tell me how to take each one to the next level, including the ones that already scored high. If something is at 9, tell me how to get to 9.5. If it is at 9.5, tell me how to get to 9.8. There is always a next level.

That last instruction turns the rubric from a verdict into a tool. You learn the smallest move that would lift the score, and that move exists at every level. The AI does not get to declare you finished; you decide when to stop. ¹

Continue to the brainstorm-iterate loop.