← Back to README · Next: Low vs high fidelity questions
What the skills do
Grill MeandGrill with Docsare used worldwide as a replacement for plan mode in agents (0:00).- Their design intent: relentlessly question you until you and the agent reach a shared understanding of the thing to be built (
0:00). - Common complaint the author winces at: "Codex just asked me 200 questions" — treated as a symptom of misuse, not a bug in the skill (
0:00).
Why the failure modes are user-side
- The skills' quality depends on the skill of the person answering (
0:25). - To use them well, you must be good at planning, which means you need:
- The skills themselves are short files — they are designed to aid an engineer, not replace one (
0:25).
Structure of the video
- Nine things people get wrong, but framed first through a set of lenses for understanding the failure modes — without the right framing you cannot change the behaviour (
0:53).
The lenses that follow:
| Lens | Failure it explains |
|---|---|
| Question fidelity | Trying to answer ungrillable questions in a grilling session |
| Scope | Hidden high-fidelity questions + hitting the model's dumb zone |
| Active vs passive | 540-question interviews, or endless grilling instead of coding |
| Context value | Throwing away the session's design decisions |
| Model choice | Using a model too weak for parametric-knowledge work |
| Parallelism | Leaving throughput on the table |