I started thinking about context with AI after a post by Antonio Leiva (opens in a new tab) about planning by features, scenarios in Gherkin and tested code. The idea was simple: if the team does not fix the context before writing code, AI fills the gaps. And when it fills the gaps, it decides for you.
AI does not fail for lack of capability, it fails for lack of explicit agreements.
From that a spec-driven flow was born: plan before you execute, persist the context and move scenario by scenario. Not as bureaucracy, but as the project's living contract. Clear goals, acceptance criteria, risks, Definition of Done and traceable progress. When all of that exists, AI stops improvising and starts collaborating.
The key is not the prompt, it is the system.
Treating AI as a member of the team means clear limits: declared skills, with what it can and cannot do; versioned planning; visible standards; and a small, curated knowledge base for RAG. Less generic documentation and more authorised sources, with an owner, a date and an expiry. If the context is not in the repository, it does not exist.
In parallel, the models themselves have started publishing guides on how to work in a team and how to manage context. Proposals like Google Antigravity have appeared too, integrating planning, todos and context directly in the IDE. For individual work it can be a very good solution. In a team I have reservations: if the context does not live in the code, it is neither shared nor audited. The skills, the scenarios and the decisions should be part of the project, not of a local view.
For that same need I am working on .ai-dev (opens in a new tab): a way to declare the context of the application explicitly so AI works aligned with the way of working, the team's standards, compliance, quality and security. It is far from finished, but the direction is clear: we need an AI architecture for software engineering. Only then does AI stop being an individual shortcut and start adding real, sustainable value to teamwork.
AI does not need to go faster, it needs to work inside the team's agreements.