Axios Finish Line: How to AI

Why it matters: Ask top large language models like ChatGPT something simple or generic, and you will get a simple or generic answer. Ask the right questions the right way, and you will often get magic.

Jim asked Dan Cox, our CTO and AI leader, to help him craft six ways for ordinary users to get more extraordinary answers. It starts with prompts — the very questions you pose to ChatGPT, Claude, Gemini, Grok or Perplexity.

• Just in time for performance-review season, here are six tips (for both reviewers & reviewees):

1. It's a conversation, not a search engine. The biggest mistake newbies make is treating AI like Google — one question, one answer, done. The magic happens in the back-and-forth. Ask a question. Read the answer. Then push: "Make it shorter ... Give me three alternatives ... That's too formal ... What am I missing?" The best outputs come from the fifth or sixth exchange.

• Example: You ask for help requesting a raise. The first draft is generic. You say: "Too corporate. I've been here six years and my boss is informal — make it sound like a real conversation." Now it's useful.

2. Nail the Who. Start by explaining who you are — your role, experience, anything relevant — and who you want the AI to think like when answering. Be specific.

• Example: "I'm a senior account manager at a midsized software company. I've been here six years, consistently hit my numbers, and just took on two new direct reports. I want you to think like a brutally honest executive coach who has helped hundreds of people negotiate compensation."

3. Context matters most. Give detailed, real-world framing upfront. The AI doesn't know your situation, your audience, or what you're trying to avoid unless you tell it. Specificity is everything.

• Example (building on the Who): "My annual review is in two weeks. I haven't had a raise in 18 months despite a promotion in title. I know the company had a rough Q3, but my division exceeded targets. My boss is supportive, but not the final decision-maker — he has to pitch it to the VP. What's the smartest approach?"

4. Just say no. Tell it what not to do. This sharpens output dramatically.

• Example (adding constraints): "Don't give me generic advice like 'know your worth.' Don't suggest ultimatums — I'm not bluffing. And don't make it sound like a script I'd read verbatim."

5. Say: "Think step by step." When you're dealing with anything complex — a negotiation, a decision with trade-offs, a strategy with multiple variables — ask the AI to reason through it explicitly. This simple phrase dramatically improves output quality.

• Example: "Think step by step about how my boss will react and what objections he might raise when pitching this to the VP."

6. Just dump the image in. The models are extraordinary at instantly understanding screenshots, documents or files. Stop wasting time explaining what you're looking at — the AI can just see it.

• Example: Screenshot your company's salary bands from the internal HR portal. Paste it into ChatGPT alongside your title and tenure. Ask: "Based on this, where should I be? What's a reasonable ask?"

⚠️ Trust but verify. AI can hallucinate confidently — inventing facts, statistics, even citations that don't exist. The more specific the claim, the more you should double-check. Use it to think, draft and strategize. But if it spits out a number or a name, verify it before you repeat it.

The bottom line: AI is a power tool. It rewards users who treat it like a sharp colleague rather than a magic box. Be specific. Be demanding. Keep pushing.

• How'd you do? How'd we do? What's your power prompt? Let us know: [email protected].

Go deeper: Jim's video, "Blunt AI advice."