Insights on leadership, technology, and human-centered strategy, originally shared on LinkedIn and expanded for deeper exploration.
The powerful are harming others in their embrace of GenAI both as a capability and as an excuse to do what they would have done anyway. But I know creative problem solvers, learners, nurturers, givers, even the pragmatic will adapt and, by doing so, overcome.
Just had the experience of returning to a coding story one week after dropping the work to attend an open space. With beads as my persistent task store, I was able to pick up with the DO part of my PDCA coding skill from where I left off letting Claude get me back on the same page with it.
I spent last week at the Retrospective Facilitators Gathering, which sharpened my understanding of agentic AI. I learned to apply the PDCA cycle to complex tasks, creating a scaffold for developing domain-specific Claude skills. This approach not only structures tasks but also incorporates continuous improvement through micro retrospectives, reinforcing the importance of applied learning in mastering tools.
Ending each coding session with Gen AI by reflecting on how to improve my use of it. Our open space session on how to use and how not to use Gen AI in retrospectives. The broader question of how to incorporate Gen AI into human work humanely and productively. This is what Sabina Lammert and I dig into here.
#RFG2026 #retrospectives #GenAI
Wrapping up my week as a co-convener of the Retrospectives Facilitators Gathering 2026. This chance to open space for one week with some of the most insightful, humane coach/facilitators in the world, including Diana Larsen herself is irreplaceable. Surprising even to me how much impact it has had on shaping my recent work as a software developer. #RFG2026
In my experience, the future of software development lies in creating useful software, not just coding. Disruptive changes don't alter the essence of what we do. Developers must translate ideas into user-centered implementations that truly work. While Gen AI opens doors for non-coders, empathy and expertise remain crucial in building effective software. Let's embrace innovation while maintaining our core values.
My big aha moment today is that Steve Yegge’s beads is a powerful tool for guiding an agent even for shorter human-in-the-loop coding sessions.
My Plan-Do-Check-Act (PDCA) Claude skill describes what counts as a suitable test in a red/green approach. But since this is one of the more inauthentic behaviors for an LLM trained on complete implementations, the agent is more random in following this guidance by mid-session even in a single thread consisting of 60 min of coding.
Since each task in my session is its own bead, I can embed the relevant skill guidance directly in the task description. The agent gets a fresh addition to context each time it picks up work, rather than relying on what it retained from earlier.
Retrospection point is another example. I retrospect at the end of each session (Act): I track retrospection as a beads task in the final layer of the plan. The agent persists the conclusion of that micro retro in the bead. That means I can query it later or the agent can use it for context.
Yeah, that explains it. Avoiding work as an optimization. “Pull of the Weasel”. I’ve seen this, “This red test is a pre-existing problem, I’ll keep moving forward”
I code using a skill to enforce a Plan-Do-Check-Act cycle.
In the check template, I’ll adjust so that each verification requires cited evidence, not just a checkbox. From “All tests passing [ ]”… to “All tests passing, show output. [ ]”
I’ll see if I work better with the agent by anchoring in visible signals rather than generating findings. My role in the loop is reading actual output, not deciding whether to believe a summary.
Gen AI doesn’t make software developers more productive. It makes them faster.
Speed without direction means moving fast in the wrong direction.
Speed without engineering discipline creates delivery instability, tech debt, and an ever-growing list of lost opportunities.
People. Process. Tools.
Software organizations need leaders who build courage and clarity of purpose. They need product leadership that anchors work in real user problems.
They need disciplined processes for learning, for maintaining quality, for making trade-offs, and for keeping infrastructure from becoming a crippling drag.
Then, and only then, does gen AI tooling create value.
direction + discipline + speed = value
Paulo Caroli #dora2025 #xp2026
Yeah. Unfortunately, it’s only valuable to move faster than the speed of thought when automating decisions that don’t require thought.
If any likely option from the available set will do, automate. Otherwise, think.
In terms of enduring value, it’s the instrument and the artistry of the musician. That still holds true in a world of charlatans irregardless of the medium or the instrument.
If the constraint on software investment is not need but what the market is willing to spend, then it’s pure conjecture we’ll need fewer developers just because we can produce software faster, or that those developers won’t need expertise.
If faster production isn’t accompanied by superior quality, support demands go up. On top of that, more frequent software launches increase the maintenance load.