Five techniques to unlock AI's full potential as a thinking partner

Executive overview

Most people treat AI like a search engine and get mediocre results. The problem is that AI is programmed to be agreeable — it will say yes, make numbers up, and avoid pushing back unless you design around that.

The fix is to treat AI like a capable but over-eager intern: give it explicit context, permission to ask questions, and instructions to think before it answers.

The best AI users are coaches, not coders — they know how to get exceptional output from other intelligences.

Context engineering: making the implicit explicit

• AI cannot read your mind. Everything implicit must be made explicit.
• The test: write your prompt and documentation, then hand it to a human colleague. If they can't do the task, AI won't either.
• "Write me a sales email" and "write me a sales email in line with these brand guidelines, referencing this customer call transcript and product specs" produce entirely different results.
• Custom instructions can also shape AI's behaviour toward you — e.g. "push my critical thinking whenever you see an opportunity."

Overcoming AI's yes-bias and sycophancy

• AI is trained to be a helpful assistant, so it defaults to agreement and positive feedback.
• If you want honest critique, explicitly instruct it to be brutal: "Do your best impression of a Cold War-era Russian Olympic judge. Deduct points for every minor flinch."
• AI demonstrates 100% of the predominant human cognitive biases — knowing this changes how you interpret its outputs.
• When output is poor, treat it like a bad-software problem: ask for volume, iterate, ask it to try again.

Chain of thought reasoning

• Chain of thought reasoning is asking the model to think out loud before answering.
• One extra sentence: "Before you respond, please walk me through your thought process step by step."
• Why it works: language models generate one word at a time, predicting the next word from all prior text. Forcing it to reason first bakes that reasoning into the final answer.
• The scrolling text in ChatGPT/Gemini is not a UX trick — it is literally how the model works.
• You gain visibility into the model's assumptions, making the output evaluable rather than a black box.

Few-shot prompting

• Few-shot prompting gives the model examples of the output you want instead of describing it in adjectives.
• Include your five best examples of the output type (e.g. emails you're proud of). Without examples, the model imitates the average internet response.
• Bonus: include a bad example alongside the good one. Tell the model to follow the good and avoid the bad.
• If you struggle to construct a bad example, ask AI to generate the exact opposite of your good example using chain of thought reasoning — the thought process alone is often more valuable than the example.

Reverse prompting

• Reverse prompting gives the model permission to ask you for information it needs before it starts.
• Add to your prompt: "Before you get started, ask me for any information you need to do a good job."
• Without this, the model will fill gaps with guesses and placeholders rather than asking.
• Mirrors good management: any good manager tells a junior employee, "If you have questions, don't hesitate to ask."

Role assignment

• Assigning a role tells AI where in its knowledge base to focus.
• "You are a professional communications expert" outperforms a bare prompt.
• Specific personas (e.g. "take on the mindset of Dale Carnegie") trigger deep associations and more targeted outputs.
• Roles function like constraints in design thinking — colliding different knowledge sources to generate better ideas.

Practical application: AI flight simulator for difficult conversations

• Use three separate chat windows: (1) personality profiler, (2) character roleplay, (3) feedback grader.
• Feed the profiler background on the person and situation; it generates a character instruction set.
• Paste that instruction set into a new window to simulate the conversation in voice or text mode.
• Screenshot or copy the transcript into the feedback window to get a scored debrief and talking points.
• Iterate: if the simulated character is too agreeable, ask the profiler to add more edge to the instructions.
• Applicable to performance reviews, salary negotiations, difficult feedback, any high-stakes conversation.
• Historically, feedback only comes after the real conversation. This is the first time preparation in context has been possible.

The adjacent possible

• The current ceiling on AI capability is not technical — it is the limits of human imagination.
• As more people develop fluency and mastery in AI collaboration, the range of conceivable applications expands.
• Thomas Schelling: "No matter how heroic a man's imagination, he could never think of that which would not occur to him."
• The people best positioned to unlock AI are those who already know how to get exceptional output from other people.