Treat AI Like Your Genius Intern

In case it’s helpful, I shared this AI cheat sheet with a bunch of exec friends today:

1. Start with a tight leash

Think of AI like a genius intern. High agency, high intelligence, and raw horsepower. The worst thing you can do when the genius intern joins your company is give them 5 tasks and say “go.” They’ll end up doing a C+ job at each task. Instead, you start with a very short leash, give them one low-stakes job, and offer enough feedback until you they’ve reached your standard of good & earned some trust. Then you extend the leash little-by-little, as quantity of feedback goes down, quality of work goes up, and trust grows.

2. Don’t half-ass the onboarding

“AI is broken. It either hallucinates or generates shitty output.” Can’t tell you how many times friends have shared this complaint with me. What they fail to mention is they gave the model shitty context, their prompt was the byproduct of lackluster thinking, and after getting their first response from a model 10 minutes in they assumed the technology was trash and didn’t push through. Going back to the intern corollary, in what world would you expect them to perform well if you don’t provide great context, if you aren’t clear with your expectations, and if you don’t give feedback until they know what good looks like?

3. Assume it’s a you issue

In 2022, GPT 3.5 could run for 36 seconds and complete a job successfully 50% of the time. Just 4 years later, Opus 4.6 can run for 10 hours and achieve the same success rate. Every 7 months, the time an AI agent can run for & achieve equivalent completion is doubling. What does this mean? With every day that passes, you can expect AI to behave far more like quality labor and less like advanced autocomplete. And you can safely assume that poor performance is more likely human skill issue and less likely faulty technology.

4. Fire yourself out of your job

The best way to AI proof yourself is to fire yourself before AI does. It’s like those Chinese finger traps. The only way to escape is to fight your instinct to resist, but instead lean in. Get into the cycle of abstraction. Identify a process. Offload it to AI chunk-by-chunk. Push to the point of discomfort, where you serve as the first mile and final mile for the AI. And then move to the next process. Force yourself to elevate to the layer of direction & taste.

5. Get to good, then codify

Bear with me on the genius intern analogy once more. Once you’ve followed rule 1 and 2, and your AI has produced an output that you consider great, codify it. In business terms, we call this an SOP. In AI, it’s a skill. Just as you use an SOP to ramp up future hires and get them to generate quality, standardized work, you use an AI skill to get a probabilistic technology to behave more deterministically.

6. Make your AI proactive

Once you’ve got the skill schtick down, it’s time to go from reactive to proactive. Create a scheduled task where Claude/ChatGPT scans your conversations and connected apps (Gmail, slack, etc) from the previous week and recommends repetitive work that should be turned into repeatable skills. Now, your work gets automated while you sleep.

7. Less reading more building

I love consuming content about AI. Karpathy and 3Blue1Brown help me understand the technology. Levie helps me apply it. But this is all secondary to being in the trenches, firing up the tools and learning the old fashioned way…by building. By struggling, failing, hitting errors, asking Claude what went wrong, and pushing through to the point of success. No better teacher.

8. Spend more time with engineers

My AI aha moment happened 18 months ago. I watched 2 ai-enabled engineers put out as much quality software as 8 peers coding the old-fashioned way. I quickly realized this technology is not evenly distributed. If you want to know what AI will be capable of in knowledge work 6 months from now, velcro yourself to your technical friends and see how they’re using the technology today.