← Previous: Low vs high fidelity questions · README · Next: Being active, not passive
Scope = how large a thing you are grilling (3:44).
Too-large scope causes two distinct problems (3:44):
Problem 1 — hidden high-fidelity questions
- A large scope conceals high-fidelity questions inside it that are hard to answer without seeing the whole thing built (
3:44). - Principle: it is always easier to build off something you know works than to endlessly plan scope out into the future (
3:44). - People who schedule days and days of tasks end up with poor results because they aren't building on a foundation they're aligned with — they pushed too far into the future without something solid underneath (
4:14).
Problem 2 — the model's "dumb zone"
- ~120k tokens is where the author estimates the dumb zone begins for most state-of-the-art models (
5:07). - Past that point the model's attention relationships get strained and it starts making stupider decisions (
5:07). - A big-scope grill starts on a near-empty context window but grows until you're in the dumb zone before you've answered half the questions — forcing an awkward handoff or compaction (
4:42). - All of that is avoidable by picking a smaller scope up front, so the whole grill fits comfortably in the smart zone (
4:42).
flowchart TD
L["Large scope<br/>(too big for one agent)"] -->|ask the agent to break it down<br/>ahead of time| S1[Small scope A]
L --> S2[Small scope B]
L --> S3[Small scope C]
S1 --> G1[Grill individually,<br/>stay under ~120k]
S2 --> G1
S3 --> G1