Intent, Context, Judgement

AI-first development (e.g via Spec-Driven-Development) is not only about powerful models. In coding, quality outputs come from three things working together: clarity of intent, effective context, and human judgement at the right time.

A lot of our recent work has focused on improving all three.

Clear intent matters because it gives the AI good steering. Coding agents need enough direction to understand what to build and what good looks like. But intent on its own is not enough. In practice, teams often spend a lot of time trying to connect prompts to the right context, or they miss that step altogether and hope the AI will discover it for itself. The result is usually avoidable review loops, rework, or solutions that look convincing but do not fit how we actually build.

Frontier models are continuously improving, and that makes it easier than ever to get high-quality outputs from relatively simple prompts. Their coding capability is remarkable, but even the best models can still hallucinate, or generate polished solutions that are not aligned with our standards, patterns, and principles.

We experiment in different ways to close the gap. One technique is building context infrastructure close to the code. This is a shared system of reusable knowledge, standards, patterns, guidance, decisions, and references that helps both humans and AI do work consistently and effectively. The aim is to make the right context available in the right format at the right time, so AI is not forced to guess (or fall-back to its general knowledge) how we work or what good looks like.

Within that, skills act as a thin steering layer between the agent and the canonical information it needs. Their job is not to contain all the knowledge themselves, but to route the agent to the right context, add decision rules, and help it act more like an engineer working within our environment. This keeps the knowledge portable, reduces duplication, and makes it easier to reuse the same context across different AI workflows.

We are also very deliberate about keeping this simple. Less is more. The value is not in creating lots of skills, but in creating the right ones, making sure they are used at the right times, and checking whether they are genuinely improving outcomes. A stale, noisy, or over-broad skill is worse than none at all.

Context also exists in layers. Some guidance belongs at team level, some at domain level, and some at enterprise level. The goal is to make that knowledge structured, portable, and usable wherever AI is helping with planning or execution.

What feels important to us is that this is a maturing discipline. We are moving beyond the idea that better AI outcomes come from better prompts alone, and treating quality as a delivery system made up of intent, context, and judgement. That feels like a more realistic and sustainable model for AI-first development.

The next step in maturity is not just about creating context and skills, but about governing them well. This means measuring what works, pruning what doesn't and treating them as real delivery assets that need ownership, review, and continuous improvement.

For us, AI-first development is not about replacing engineering judgement. It is about creating the conditions for better AI decisions - clearer intent, stronger context, and human judgement applied where it matters most.