Working memory holds about four items. Each layer of recursion consumes one slot just to hold the frame. Three levels deep and you’re reasoning. Four and you’re juggling.
Watch it happen in process design:
- Write a document.
- Decide which document to write.
- Design the system that decides which documents to write.
- Improve the system that designs the system that decides.
Level 1 is execution. Level 2 is management. Level 3 is strategy. Level 4 is where most people’s eyes glaze over.
We can’t think in four dimensions either. But we have a trick for that: projection. We slice 4D objects into 3D cross-sections and reason about those instead. Nobody visualises a tesseract directly. They rotate it and watch the shadows.
Skills are the same trick applied to process recursion.
A well-written skill — a checklist, a framework, a heuristic, a rule of thumb — is a 3D projection of a 4D insight. It collapses one recursion level into a flat instruction so your working memory doesn’t have to hold the frame.
“Always test as the user, not just smoke test” is a collapsed version of a painful debugging session, a root cause analysis, a system design review, and a meta-observation about what kinds of testing prevent what kinds of failures. The recursion is four levels deep. The skill is one sentence.
This is why experienced practitioners seem to operate effortlessly at high levels of abstraction. They’re not smarter. They have more collapsed layers. Each rule they’ve internalised frees a working memory slot for the next level up.
It’s also why teaching is hard. You can hand someone the 3D projection, but they can’t unfold it back into 4D without doing the recursion themselves. The rule makes sense. The judgment behind the rule — knowing when to break it — requires having climbed the stack at least once.
The implication for AI systems: if you can externalise your collapsed recursion into a searchable, structured store, you’ve done something humans can’t. A skill file doesn’t forget. It doesn’t consume working memory. And it’s legible to agents who can apply it at scale without ever having climbed the stack themselves.
They’re running on your projections.