I have been bullish on AI for a while, and for good reason. Coming from mainframe-era software delivery in regulated industries, I have seen how advances in data science and machine learning can fundamentally change what is possible in IT.
What feels different now is the speed and quality at which we can turn data into actionable knowledge. That shift has opened new possibilities for organisations, teams, and individual developers.
The Lesson I Learned the Hard Way
I have leaned heavily into AI assistants such as Cursor and Claude, and their capability has improved dramatically. But the biggest unlock was not model quality alone. It was communication quality.
As a colleague once told me, I often spoke as if everyone could read my mind. AI tooling quickly exposed the cost of that habit. Vague asks produce vague outputs. Structure produces leverage.
AI Assistants Are a Team, Not a Toy
You would never hand a development team an unclear brief and expect production-grade output. The same rule applies here. Treat assistants as collaborative engineers: define intent, constraints, architecture boundaries, and quality standards up front.
For example, "build a data connector API" is too broad. A useful brief includes source and target systems, endpoint requirements, auth model, performance expectations, error handling, and security constraints.
Practical Rules That Work
- Plan first: define the system shape and decompose work into modular components.
- Communicate clearly: state intent, constraints, non-goals, and priorities explicitly.
- Iterate thoughtfully: review generated code critically, then refine prompts and architecture decisions in loops.
AI-assisted coding is a genuine step change, but it still rewards the fundamentals. Grease the wheels with clarity and structure, and these tools become far more than autocomplete. The teams that extract the most value are not the ones with the most advanced tools — they are the ones with the clearest thinking. Ambiguity that would cost an hour with a human developer costs nothing with AI because the output simply absorbs it and moves on. You only discover the price when you try to use what you built.