How to Steer AI When It Runs With the Wrong Thing

You ask AI for something, the answer comes back close, and then it locks onto that first version and builds everything else on top of it. This week's skill saves the most time: steering. You catch the drift early and pull it back, keep what's working, and fix the rest. We've spent a few weeks now on getting good at AI, what strong output looks like, and what to leave out of a prompt. Steering is what you do once the answer's back. Here's how it played out for me in a recent engagement.

I recently built onboarding documents for a client. I had AI pull from the principal's past workflows, their calls, and how their current onboarding actually runs. What came back was close. It also locked onto one thing: a misstep with a single customer in the current process, and it built the whole document around fixing that one incident. The real problem was bigger than that. This client had no structure or accountability across onboarding, and the customer misstep was a symptom.

I caught it fast here because I already knew the real problem, but the tell is in the shape of the draft: when AI hangs an entire document on one example, that's usually the drift, and you can spot it before you know the fix yourself.

So I kept the draft and steered it. I told it something close to this: "That customer issue is one instance of a bigger gap. The real problem is that this client has no structure and no accountability across onboarding. Step back and map the whole thing before you rewrite." The process map it had built the first time was solid, so I kept that too. A few sentences of redirect saved me the rebuild.

I turned that redirect into a reusable prompt: a 5-step steering sequence you paste the moment a draft comes back close but off, so it keeps what's working and re-scopes the rest with no rebuild. It's yours as a CoS Signal member, along with the full prompt library. Founding rate is locked at $14.99/month or $99/year.

The research backs the hard part of this. In a 2024 experimental study, researchers tested GPT-4, GPT-4o, and GPT-3.5 on anchoring bias, the tendency to let the first piece of information pull every answer that follows toward it. The more capable models were the most consistently anchored. When the researchers told the models to ignore the anchor, the bias held. The correction that helped was feeding the model a fuller set of angles so it stopped fixating on the one piece. Steering a draft is how you apply that after the answer's already back. You widen what the model is working from, and the old scope no longer fits. For a CoS handing work to AI all day, that's the habit worth building, and it takes seconds once you learn to spot the fixation. (Anchoring Bias in Large Language Models: An Experimental Study, 2024)

Heads up, this one’s an ad.

Tired of overpaying GoDaddy or Namecheap? There's a smarter option with Porkbun!

Porkbun is the domain registrar trusted by creators, developers, entrepreneurs, and anyone who wants low prices without the nonsense or the forced upsells.

Here's why so many people are making the move to the Bun:
• Most domains sold at cost
• Low, transparent registration and renewal pricing
• Free features like WHOIS privacy and SSL certificates
• Powerful web and email hosting options
• Real human support 24/7, 365 days a year
• Named the #1 domain registrar by Forbes Advisor and USA Today

Whether you're launching a business, building a personal brand, starting a side project, or creating your very first website, Porkbun makes the whole thing easy.

Get $1 off your next domain registration with Porkbun now.

Get Your Domain At Porkbun

Best practices for prompt engineering for 2026 (Anthropic) is a clear, free walkthrough of how to set up a task so AI gets closer on the first pass, which means less steering later.

Here's the shift worth keeping: most of getting good at AI happens after the first answer lands. The skill is reading what's off and steering it in place, and it's faster than starting fresh every time. Next week we turn these one-off saves into a system you can run without thinking about it. If you do one thing this week, catch the next answer that's mostly right and slightly off before you accept it, name what it got wrong about the scope, and steer it there. See how much of the work you get to keep.