LLMs are getting smaller and more efficient.
Which raises an interesting question:
Will infrastructure become a product capability again?
Today, most teams rely on APIs (OpenAI, Anthropic, etc.).
But smaller models are getting good enough to run locally.
That shifts the problem:
→ less prompting
→ more system design & infra decisions
AI becomes less a tool and more a system.
Curious:
Are you already running models locally?
A question that has been on my mind this week:
Will ops thinking become more important as vibe coding becomes common?
I’m noticing that AI-assisted development lowers the barrier to building. Even when I don’t yet know exactly how to solve a problem, I can describe my intent, iterate quickly, and generate large parts of the functionality.
It works surprisingly well but when speed shifts what you need to pay attention to.
A few practices from me that seem important:
– Reviewing AI-generated changes (even if only summaries)
– Introducing structure early: version control, CI pipelines, automated tests
– Thinking about system design before prompting
Vibe coding is fantastic for exploring solution spaces quickly.
But without some operational discipline, it’s easy to build something that works today and breaks tomorrow.
Curious how others balance speed and stability in these workflows.
Eine Sache, die mir in der AI-assisted Entwicklung auffällt:
Der Rhythmus verändert sich. Discovery wird schneller, während sich Umsetzungen verzögern können.
Der Fokus verschiebt sich damit an Stellen, an welchen Klarheit erforderlich ist.
New article available on LinkedIn:
I’ve been thinking about a subtle effect of AI-assisted product work:
Output starts to feel like progress.
You can generate structured documents, working code, or polished slides in minutes.
But ambiguity doesn’t disappear just because execution becomes cheap.
If the problem is unclear, AI will still produce something impressive-looking.
Which makes focus and critical thinking even more important.
The constraint is no longer production capacity. It’s focus, clarity and ownership.
In my workflow now I pick models by task: Opus for big planning and analysis, GPT-5.3-Codex when it’s time to build and iterate.
Planning and strategy with one and implementation with the other.
It really is a lot of fun!!!
Lately I’m questioning how long “professional prompting” will matter.
With frameworks like BMAD v6, I hardly prompt anymore.
I initiate workflows, provide context, answer questions.
The agents handle the prompting.
The work shifts from phrasing to intent and decisions.
Feels like an early but important signal for me.
GPT-5.3-Codex pushes coding a bit further into the background.
Reasoning feels stronger, feedback loops tighter.
For PMs, that shifts the work again.
Execution gets cheaper. Judgment gets louder.
I’m genuinely curious how this will play out in BMAD v6 — not to ship faster, but to think clearer and, hopefully, build better products.
BMAD v6 update: planning sessions.
Instead of prompting the AI directly, you start by loading a workflow with explicit steps, roles, and constraints.
The prompt itself feels like a meeting agenda.
Sometimes it’s close to rubber ducking, but with a useful side effect:
vague ideas are forced to become explicit decisions.
This line matters:
“IT IS CRITICAL THAT YOU FOLLOW THESE STEPS”
BMAD doesn’t remove rigor.
It anchors rigor in planning before execution becomes cheap.
Still early, but this already changes how vibe coding feels.
