stop fighting the AI for the keyboard
I would actively go out of my way to hide the fact that I used AI to code my bot. It all feels a little embarrassing thinking about it now, but at the time you really felt like a fraud from all the online discourse.
Then Peter Steinberger vibe-coded OpenClaw. He built one of the most successful agents out there, proving you can make really amazing projects completely with AI.
But you can’t let it be “Jesus take the wheel,” or else you end up with slop And if we hold on too tightly, we leave all effectiveness gains on the table.
There’s got to be of course a better way
But letting go of the keyboard is terrifying. It feels like a dog’s age since I learned Python just to build my bot, spending hours digging through the API just to get it to attack. Giving up that hard-earned control felt wrong.
Earlier attempts at letting AI build my bot landed in disaster. They’ve gotten WAY better since then, but I was still terrified they would just run amok in my codebase. I knew I could build so much more if I used it to its full potential, but it wasn’t until I listened to that Lex Fridman podcast breaking down creation of OpenClaw that I decided it was time to give real agentic coding a legit go.
I set up an AGENTS.md file and spun up an entirely new bot from scratch so I wouldn’t accidentally nuke my main projects, which gave me a safe space to finally experiment.
What I figured out pretty quickly is that AI is incredible at fixing discrete problems but completely plays it safe when it comes to design or creativity. It really only knows what it knows. If you give it examples and suggestions it can explore further, but it absolutely needs your push to get there.
The trick is giving the AI enough room to play while staying in the loop so you know exactly where it’s taking you. You stop being just a coder and become a code director, providing the overall vision and the boundaries it needs to operate inside.
You honestly don’t need to know every single line of code anymore. As long as you understand the architecture and review the AI’s work, you can steer the ship. It’s exactly like being a captain on a boat. You can’t physically do every job yourself, but if you listen to feedback from the crew, you can give the exact orders needed to reach your destination.
⌨️ Next Commit
This week, add a rule to your AGENTS.md file. We’re calling it a Shadow Review.
Tell the AI to end every task with three bullets:
- Biggest assumption
- Most likely failure
- Smallest robustness fix, limited to 10 lines of code
Review those three bullets before accepting the AI’s work. Your judgment is the final step.
May the Bugs Be Ever In your Favour🪲
Email Preference:
Unsubscribe | Update your profile | 600 1st Ave, Ste 330 PMB 92768, Seattle, WA 98104-2246