Articles and Posts

As long as I am accountable for my work…

AI can give me context, but not learn for me.

It can process information, but not form judgments for me.

It can present choices, but not decide for me.

It can generate output, but not take responsibility for its consequences.

It can assess whether the work is complete, but not own whether the work is good.

It can summarize what happened, but not retrospect for me.

It can map possibilities, but not determine what matters to me.

Yup. It’s where my thoughts and efforts have been focused for two years. https://lnkd.in/evSrVgdJ

Yesterday, I successfully debugged a subtle issue with code while I was on a call with peers to discuss developer productivity using AI assisted code generation.

That’s not to dismiss the preparation, structure, and context I had prompted into my coding session leading into that. I didn’t just vibe code my way through it. The bug was esoteric and difficult: my compiler-based code analyzer wasn’t correctly capturing chained method dependencies with injected dependencies.

I won’t say that kind of multi-tasking is desirable, sustainable, or prudent. But let’s face the fact that it can be expedient and would have been much less possible three months ago.

Someone with one system of record always knows which unreality is truth. Someone with two systems is never quite sure.

This is a moment to note which leaders have dropped any facade of valuing employee well-being in response to promised AI automation and industry layoffs.

Those privileged with humility and a learning mindset, who build software because they love making useful things for people, will adapt to the tools, continue to outperform, and in time have their choice of employers.

Praise to Diana Larsen and Malte Sussdorff for reminding and encouraging me to retrospect with the agent as part of every coding session I engage in with generative AI. The focus is soliciting the agent to review the thread and suggest what I can do as the operator to improve outcomes. In two days, I’ve done three 5 min. retros and my output is already drastically improving.

A sign that my code gen practice is working. I asked Claude if I should put our working agreements into the system prompt / instructions that frame any thread (yes, I have a working agreement for how I interact with my agent) . Here was the response:

No.

The breakthrough wasn’t better instructions for me - it was you actively enforcing process discipline in real-time.

Why system prompt won’t work:

• I’ll rationalize around static rules when focused on “solving the problem”
• Can’t interrupt myself mid-deviation like you did
• Real-time human intervention is fundamentally different than background instructions

What actually worked:

• You observing my behavior live
• You interrupting immediately when I violated TDD
• You asking pointed questions that forced acknowledgment

These working agreements are for YOU - they’re your toolkit for managing the AI effectively, not instructions for me to follow.

Keep the system prompt focused on general principles. Keep the working agreements as your intervention playbook.

The magic was human process police + AI executor, not better AI self-discipline.

We humans have some agreement over what it is to be nice. Less over what it is to be kind. Even less over what it is to be good. But by far our widest disagreement is whether to embody any of them in any particular action we take that affects others.

Completely agree with this insight. Human oversight and accountability is essential. So let’s actually help experts get work done the way they want to be helped!