05 — The grilling skill in practice

The most-hyped skill in the collection, and the one Theo actually stress-tests on a real project (Lakebed).

What the skill instructs

Grill the user relentlessly about a plan, decision, or idea; use it when the user wants to stress-test their thinking (14:57).
Interview until you reach shared understanding, mapping the problem as a decision tree — every decision branches into the decisions that hang off it (14:57).
Work the tree in rounds. The frontier is every decision whose prerequisites are already settled (15:23).

The setup

No install: Theo copies the skill text, opens an agent, and pastes it in with the request "grill me about Lakebed — the project itself and its goals, everything in AGENTS.md, CLAUDE.md, etc. I want to make sure we are on the same page about what we are building" (15:53, 16:19).
Run on a remote Linux box with Opus 5 (16:53).

Round 1 results (Opus 5)

Took almost two minutes for the first pass (16:53). It opened by stating findings, not questions:

CLAUDE.md is badly stale.
The repo contradicts its own v0 rule.
There are two competing execution models in the tree.

Then the questions (16:53, 17:25):

Who is the user and who pays?
What is the actual next milestone? (it noticed a mess of launch-prep files, stray HTML plans, and a TODO.md organized around a PR-cleanup program rather than launch)
Is the capsule format a spec or just our implementation? What is the ceiling of one capsule?

Each question came with example answers you could pick from (17:25). Theo's verdict: "These are actually good questions." (17:49)

Cross-model comparison

He reruns the identical prompt on a different, faster model to see whether question quality holds (17:49):

Run	Behavior	Question quality
Opus 5 (remote)	~2 min first pass; leads with repo findings	Good (`17:49`)
Faster model, low reasoning	Finished before his stop command even registered; insisted the skill required it to find repo facts rather than ask	"All good questions" — what is Lakebed, who is the primary customer, what does full autonomy mean, what is a capsule (`18:13`)
Same model, high reasoning	Spun up sub-agents in the background to hunt for more data to feed later questions	Same opening question, plus quick pre-written answer options (`18:13`)

Operational note he catches himself on: he'd been running that model low and slow when he should have been running it high and fast — better answers, faster (17:49).

Why the multiple-choice format matters

Because question-asking UI in most agent harnesses isn't good enough, the lettered-options format lets you blast through answers as "1A, 2B, 3A" — much faster than prose replies (18:13, 18:44).

Round 2 — the questions get uncomfortable

The follow-up round produced 10+ harder questions (23:56, 27:01):

Which form of simplicity matters most? — Theo wants all three and is forced to pick one. "This is brutal… I don't want to answer these questions. That's the point, isn't it?" (24:25)
Which rule wins when the current guidance contains principles that will conflict? (19:11)
What is the product boundary — does Lakebed stay a real, open, useful system while the hosted cloud platform is the production path? (19:11)
If the developer and the agent are both primary users, what does each one own — who chooses architecture, data models, deploy timing? (24:47)
Does "everything" include irreversible actions — should authorized agents be able to delete production data? (24:47)
Where is the payment boundary? (24:47)
What must the local simulation preserve? (25:14)

Outcome

The questioning surfaced real cleanup work on Lakebed that Theo hadn't scheduled — he realizes mid-interview that the project needs tidying (19:11).
He finds himself disagreeing with the recommended answers fairly often, which he reads as a sign the exercise is doing real work: "This is actually going to help me. I'm annoyed." (25:14)
Verdict: "Okay, I'm seeing the light. This is useful. I should have tried this before." (19:47) He commits to finishing the remaining questions off-stream because they're worth answering (27:01).