Quick Reads (Under 5 Minutes)

I’m participating in a panel at The AI Summit New York on December 11th.

What I want to discuss is using success stories, shared learnings and data to move past AI pilots that fail to pragmatic tools that deliver measurable value in people’s daily work.

As technologists and leaders, we need to talk about the discipline required to integrate LLMs as useful components of real systems. About the danger of measuring usage or time saved out of context of some meaningful outcome.

Adoption is a human activity. Return on investment results from an entire system of people, process, and tools. What’s hard about change doesn’t change just because automation is part of the solution.

“Cross-Pollinating GenAI: Transcending Departments” Dec 11, 2025  2:20 PM - 3:00 PM https://lnkd.in/gr2Jr6Xc

#TheAISummit

Claude code, perhaps because of its instructions and execution logic to reduce token usage, declares premature victory. You need to have transparency into its plan so you can interrogate any claim of completeness 🎉. “What about…” Also, check back on the goal. “Does this…”

Want to see an alternative to vibe coding?

I’m demo’ing my PDCA coding practice with LLM coding agents today at the online The State of Sustainability in Agile conference at around 12:35 pm ET (conference starts at 11am ET).

My InfoQ article and Agile Alliance blog post on the practice are linked in my featured articles list https://lnkd.in/effgt9zm

Conference info: https://lnkd.in/eFDcbJGf

Invest craft into what you do. Try to achieve some positive impact with it. Be mindful of the harm. Seek out people who do the same regardless of what they do.

An imperfect but achievable path to meaningful work.

The time investment in building an AI agent isn’t in the initial coding, it’s in developing the context, persona, and instructions for it to respond well. While there’s technical craft and ongoing effort in structuring that data, the important work is SME time: people who know what an authentic and useful conversation actually looks like.

They build examples of good behavior and responses, critique the persona, and create rubrics for what makes a conversation good or poor. This involves both upfront prep and ongoing review—watching agent conversations and collaborating with engineers building the system.

AI agents aren’t for rote automation. If we can provide prescriptive instructions and get a predictable outcome, we shouldn’t be using something like a large language model. We should use some other form of machine learning or a programmatic routine.

AI agents help with complex, adaptive problem solving. Their unpredictability, their tendency to go off script for good or ill, is exactly what makes them useful.

Given that, the iterative cycle I use for coding applies to any complex project using them.

Analyze the problem, plan the work, execute in small verifiable steps, review if it is complete. Finally, retrospect. Learn what to do better next time.

Engage in the details. Be ready to improvise. Intervene when things go off the rails. Judge if the work is of quality and will actually do what you need.

It’s empirical process control with feedback and adaptation. If you aren’t learning, you’re doing it wrong.

Our daughter is off to Ireland for a year hopefully to find work and more of her self. The best things - the things you wish for at night before you fall asleep - the things you work decades to make possible - arrive with both joy and sadness. This is life.

Forgot how expensive gpt-5 reasoning models are. My learning today is don’t let someone stick you with a reasoning model for an agent framework implementation until standard models don’t cut it. langgraph with gpt-5-mini, 30s response. With gpt-4o-mini, 9s response. Figuratively , lost in its own “thoughts”.

In switching to claude code, I’ve started asking the agent to write out its plan to a markdown file so that it persists. Then when I ask it to do a completeness check we have the same information and I ask it to plan out the gaps it finds as a second change document. Sounds heavyweight but creates explainability which is potentially worth the inference costs and it benefits from the detail and doesn’t get bored writing or reading it.

My writeup on my plan-do-check-act AI coding practice is live on InfoQ.

It works well with plan mode. In fact, it anticipated it.

And retrospection is alive and well!

I provide links to prompts and code metrics github actions.

Claude Code is honoring my Analysis/Plan and TDD prompts. Anecdotally, it appears to make more atomic changes than other agents. But that may be the nature of what I’m doing with it.

My claude md summarizes the intent of my PDCA cycle. Then I provide a path to my prompt at the point I want its detailed guidance. In Plan mode, I instruct it to analyze and plan in separate steps so that I can guide them separately.

I’ve put my prompts here https://lnkd.in/eddr3qtg