Smart-vs-Dumb Zones¶
Framing popularized by Dex Horthy (Human Layer) and taught by matt-pocock at AI Engineer 2026: every LLM session has a smart zone at low context occupancy and a dumb zone as context fills.
"When you're first kind of like working with an LLM … that's when the LLM is going to do its best work. … By around sort of 40% or around 100K … it starts to just get dumber. So as you continually keep adding stuff to the same context window, it just gets dumber and dumber until it's making kind of stupid decisions." — Pocock, 7:56
Mechanism¶
Attention scales quadratically in token count (analogy: adding a team to a football league adds games with every prior team). More tokens ⇒ more attention relationships ⇒ degraded precision. Context-window size (200K vs 1M) is irrelevant — the degradation inflection sits around ~100K regardless of the advertised window.
Practical rules Pocock derives¶
- Keep the system prompt tiny. "If you have 250K tokens, like I have seen people put in there, then you're just going to go straight into the dumb zone without even being able to do anything." (8:02–8:15)
- Size tasks to stay in the smart zone — old advice (Fowler refactoring, Pragmatic Programmer): don't bite off more than you can chew. Split big tasks into small ones that fit.
- Watch the token counter on every turn (Claude Code status line) — "absolutely essential."
- For review work, clear context first so the reviewer runs in the smart zone; otherwise it reviews in the dumb zone where the implementation already landed.
Related wiki¶
- memento-principle — the companion constraint (reset over compact)
- context-compaction — the technique Pocock rejects in favour of clearing
- context-engineering · claude-code-master-loop · agent-smell