Blog

Where to Start With Knowledge for AI

Part 7 of Knowledge Up, Context Down: putting it together. A self-check, a minimal setup per level, and a look ahead.

November 18, 2026
5 minutes
‌
Example scores from 1 to 4 for each of the four levels. A person starts at the level where the score is low and the pain is high.

New to this series? Start with the overview. Previous: part 6.

The previous parts describe four levels (personal, project, team and organization) and how knowledge moves between them. This part answers the practical question: where do you start? My suggestion is to start small, at the level where problems show up most. I've kept this part deliberately short. It's not a full plan for all four levels, just enough to make a start.

Self-check: where to start

Score each level, and the flow between them, from 1 to 4 with this scale:

  • Score 1
    • A level: nothing written down; knowledge lives in heads and chats
    • The flow: nothing moves between levels
  • Score 2
    • A level: some notes exist, but they're scattered and outdated
    • The flow: sharing happens now and then, with no agreed way
  • Score 3
    • A level: one clear place, an owner, and a regular update
    • The flow: an agreed way to propose and decide
  • Score 4
    • A level: AI reads it, and people can see and correct what AI learned
    • The flow: outdated knowledge also gets removed

Then write down your scores, and the biggest problem at each level: what costs you the most time there, for example decisions being reopened, duplicate work, or agents giving conflicting answers.

Start where the score is low and the problems are biggest. A low score without problems can wait.

This is simpler than existing maturity models, such as Dubois's eight levels. Those are useful for planning a larger program. This one is meant for deciding the next step.

Start with one thing to set up and one habit per level

Once you know where to start, this is the minimum for that level: one thing to set up, and one habit that keeps it current.

  • Personal
    • Set up: a profile file your AI reads first
    • Habit: a short, regular review
  • Project
    • Set up: a state file
    • Habit: record each decision, its reason and who made it
  • Team
    • Set up: one shared space everyone can see
    • Habit: a regular update by the steward
  • Organization
    • Set up: a short, versioned context file, or a curated space for business teams
    • Habit: a regular review by a named owner
  • Flow between levels
    • Set up: the signals from Part 6 – The flow
    • Habit: share useful things as a proposal, and archive what's outdated

An even smaller start

If even that feels like too much, begin with one thing that is already useful to you: a skill, a prompt, or part of your agent setup. Share it with your team, in the repository, and propose that others use it.

Then see what happens. Does the team adopt it? Do people improve it? Does someone ask where the other skills are? That shows you where your team stands in practice, and it gives you a first shared piece to build on.

For an organization, the first step is similar. Back one team that shares its first skill, with time and attention. Next to that, start a short, versioned context file, or a curated space for business teams, with the vocabulary and policies every agent needs, and give it a named owner who reviews it regularly. Running the self-check with a few teams helps you pick which one goes first.

Looking ahead

AI will take over more of the knowledge management work described in these posts: capturing decisions, summarizing, linking related knowledge, spotting conflicts, flagging what's outdated, and proposing what to move up or remove. Some of that already happens. kmctl's audit agent, for example, reports pages that no longer match the code.

Approval will shift too. Where AI proves reliable for a type of change, such as fixing a broken link or archiving a page nobody uses, people will stop checking every case. That's fine, as long as the choice to trust it is made explicitly and can be taken back.

What I expect to stay with people:

  • Intent: what you're trying to build, and why. As Part 1 – Fast with AI showed, the AI doesn't know what it doesn't know.
  • Trade-offs: priorities, what's out of scope, and which risks you accept.
  • Accountability: who answers for the official version of something. "The AI decided" is not an answer an auditor or a customer will accept.

Summary

Knowledge up, context down, in one paragraph: keep your personal context small and share what's useful early. Record project decisions and who made them while you work. Give team knowledge a steward and keep it visible. Curate a short base context for the whole organization. And move knowledge between levels when there are clear signals that others need it. That is what keeps working with AI fast, and under your control.

Open questions

  • Where did your self-check score lowest, and what will you set up first?
  • Which signals and principles don't fit your team?

I'd like to hear what you'd change.

For leaders: support the first shared skill with time and attention.

For developers: share one useful thing with your team this week.

Further reading for leaders and developers:

Back to the start: Knowledge Up, Context Down, the series overview

Contact

Let’s discuss how we can support your journey.

‌
‌
‌
‌
‌
‌
‌
‌
‌