All articles

    Architecture note

    27 March 2026by Gustavo Teixeira Cuco2 min read

    AI Context Governance

    Architecture context

    Reusable method

    Information and context

    Applied in

    Agents & SkillsLusiberiaStaysTHE iCON
    View on LinkedIn

    Most teams solve the problem of creating context for AI agents.

    Almost none solve the problem of keeping it current.

    I spent today reading two posts that, together, define the second AI problem — the one that comes after adoption.

    Marc Baselga shared how his team at Supra successfully rolled out Claude Code — agents in production, real value delivered. Then hit something unexpected: everyone had their own context files, their own versions of how the company worked, and no system connecting any of it. When something changed, no one knew which file to update. He called it an "unexpected tax."

    Leah Tharin put the philosophy behind it: generic context is generated. Correct context is earned. The value of a knowledge worker in an AI world is the delta between what the model infers and what you carefully curate as its memory.

    Both posts pointed at the same gap. Here's what I built.

    I added a trigger matrix to my base template — a table that maps every type of change (new route, schema update, new dependency, business model shift) to the exact doc that must be updated. Not as a follow-up. In the same PR. The task is not complete until the docs reflect reality.

    One line does the work: documentation maintenance is enforcement, not suggestion.

    What I didn't expect: propagating this update to 7 repos immediately exposed another gap — no convention for distinguishing decisions inherited from the base template versus decisions made locally within each repo. The system forced me to stop, fix the source of truth first, and only then propagate.

    That's the thing about governance. You don't discover the gaps when you build it. You discover them when you use it.

    The goal isn't a perfect system on day one. It's a system that makes its own gaps visible before they become problems.

    Template is open source — link in the comments. If you're building something similar or have questions about how this works in practice, drop a comment. Happy to discuss.

    Sources