AI Working Framework

Keep the human in the work

This is a working set of principles for staying thoughtful while using AI intensively. It began with two simple observations: I was opening too many threads, and I needed time away from electronics to sort things out in my own head.

Then a conversation about the framework demonstrated the problem. I brought an intuition; AI turned it into an extensive, articulate explanation. It answered the question, but I couldn't absorb the whole answer or pass the insight to a friend who hadn't shared the context. Useful output had arrived faster than useful understanding.

These are personal working rules and questions to explore, rather than a finished methodology or a claim about what works for everyone.

01 · Bound active threads

My starting limit is three or four cognitively active threads. An agent running a task still leaves me with assumptions to remember, decisions to make, and a result to review. Opening more sessions doesn't create more attention.

New ideas can go into a parking lot. They become active when another thread closes or is deliberately paused. The number is a practical starting point to test, not a universal threshold.

02 · Protect offline consolidation

Leave deliberate periods without a laptop, phone, or another stream of input. The purpose is to let the day's material settle into a model I can use: what matters, what changed, and what I should do next.

I want to resist filling every waiting moment with more output. A walk or quiet time outside can belong to the work even when it doesn't produce a visible artifact.

03 · Control the pace of explanation

When exploring an important idea, take one conceptual layer at a time. Work through an example, connect it to something familiar, disagree with it if necessary, and only then expand the map.

A compressed answer can be a useful reference after understanding has formed. It cannot always replace the process that creates that understanding. The question is how much of the explanation I can actually use.

A loop I want to practise

Choose a bounded question. Let AI help investigate or build. Review the evidence. Step away and reconstruct the idea in my own words. Write down what I now understand and the next decision that follows from it.

This is also a test for sharing an insight. Can I give someone the starting intuition, a concrete example, and a conclusion they can reason about? Passing them the entire transcript may transfer the text while losing the understanding.

Open questions

How do I notice when review becomes the bottleneck? How much practice should I preserve by doing work myself? What makes an agent handoff easy to resume? How do I distinguish useful parallelism from a growing queue of unfinished decisions?

The article series below develops these questions individually. It is deliberately unfinished: the framework should change as the practice teaches me something.