Read from left to right: the model is the same for everyone. What changes the result is the context that only you have — and grill-me exists to draw it out.
Trail map
Detailed content
🧠 Why Context Extraction Is the Whole Game
The conceptual foundation: why the model has become a commodity, what tacit knowledge is, why a brain dump fails, and what makes grill-me worthwhile.
If you and everyone else use the same Claude Opus 4.8, the same prompt produces the same result. The model is an identical starting point.
It dispels the illusion that “improving the prompt” is the differentiator. It isn’t—the differentiator is the context.
Model commoditization · same input → same output · where the real advantage lies.
When you add your context, preferences, and decisions to the model, the output starts to sound like you.
That’s exactly what skills and docs carry. Without your context, everything stays generic.
Context = advantage · taste/voice/decisions · generic vs. original.
The hard part isn’t the model understanding—it’s getting knowledge out of your head, where it’s tacit and disorganized.
Recognizing this changes the focus: the bottleneck is extraction, not prompt engineering.
Tacit vs. explicit · the extraction bottleneck · why it’s uncomfortable.
Dumping everything into Claude for 5 minutes seems sufficient, but it never is: you generalize, forget, and leave gaps.
It’s the rationale for the interview: only targeted questions reveal what a free-form dump hides.
Invisible gaps · expert bias · asking > dumping.
"If I had 6 hours to chop down a tree, I’d spend 4 sharpening the axe." Grilling is sharpening before you chop.
Justifies spending “boring” time at the start—it pays off many times over later.
Cost upfront · downstream return · productive patience.
Without grill-me, you start iteration 1 at nearly 70%. With it, nearly 90%—and you still iterate, just from a much better starting point.
Show the concrete improvement: fewer iterations to get the skill to a reliable state.
Starting point · number of iterations · never reaches 100% (and that's okay).
A markdown file in brainstorms/ where each answer is recorded. The file, not the model’s context, is the source of truth.
The context window fills up and the model forgets. The file preserves everything that was said.
Disk > model memory · checkpoint · a predictable home.
Use it to stress-test a plan, do discovery, or extract a process. Don’t use it for a trivial, well-defined task.
Knowing the limit keeps you from turning everything into an interrogation when there's nothing to extract.
Good for: design, discovery, process · bad for: obvious tasks.