🔬 Grill-me & adversarial skills

A grill-me is a skill that does just one thing: transforms the model in a adversarial interviewer. Instead of you giving an order and the AI obeying, the relationship is reversed: it asks, you answer. It asks questions, raises ideas you hadn’t considered, points out dependencies between decisions — and keeps going until the two of you reach a shared understanding about what will be done.

Pocock describes this skill as "unreasonably effective" — absurdly effective for something so small. The reason is simple: modern models tend to be too agreeable. If you say "my idea is X," AI tends to praise it and move forward, even if X has a huge hole. grill-me deliberately breaks that habit: it tells AI to disagree, find what’s missing, and challenge you. Common mistake: thinking this is just "the AI asking questions." The difference is in the stance — adversarial, not sycophantic — and the discipline of asking one question at a time (you'll see why in topic 3).

Many agents have a plan mode: the AI reads your request, writes a plan, and waits for you to approve it before touching the codebase. It’s useful, but it has a flaw: the plan is a monologue. The AI guesses what you meant and spits out a document. If it misunderstood one detail at the beginning, the whole plan goes off track — and you only notice when you reread everything.

Pocock uses grill-me in place of plan mode. His phrase sums it up well: "here's my idea, interview me, let's reach shared understanding, flush out weirdness before we implement" — here’s my idea; interview me, let’s reach a shared understanding, let’s tease out the oddities before implementing. The key difference: plan mode is AI assuming; grill-me is the AI asking. Instead of a plan that might be wrong, you build understanding together through dialogue. The reason is that misunderstandings are expensive: a wrong assumption becomes wrong code, which becomes rework.

The grill-me fits in 4–5 sentences. Every sentence is a deliberate rule — it’s not filler text, it’s dense instruction. Notice how each part does a job: "interview me relentlessly... until we reach a shared understanding" defines the goal (keep interviewing until there’s alignment). "Walk down each branch of the design tree, resolving dependencies between decisions one-by-one" gives the method: go through the design tree branch by branch. "For each question, provide your recommended answer" requires AI not only to ask, but also to recommend her answer (to give you a starting point).

The last two sentences are the secret sauce. "Ask the questions one at a time" — one question at a time. Without this, AI dumps 12 questions on you at once and you get lost; one at a time keeps you focused and lets each answer shape the next question. And "if a question can be answered by exploring the codebase, explore the codebase instead" — if the AI can answer by looking at the code, it looks at the code instead of asking you. This keeps it from grilling you about things it can figure out on its own. The common mistake here, “improving” the skill means bloating the text: the more you add, the more diluted the behavior gets. The strength lies in being concise. Below, the text exact to copy:

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer. Ask the questions one at a time. If a question can be answered by exploring the codebase, explore the codebase instead.

Free translation: "Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the decision tree, resolving dependencies between decisions one by one. For each question, give your recommended answer. Ask the questions one at a time. If a question can be answered by exploring the codebase, explore the codebase instead of asking."

The main benefit of grill-me is moving alignment to before of implementation. Pocock's phrase— "flush out weirdness before we implement" — it’s literally this: squeezing out the quirks before of implementing. Every misunderstood decision that turns into code is expensive: the AI writes it, you review it, realize it wasn't quite right, and ask it to redo it. Each of these loops burns time and tokens. The grill-me cuts these loops off at the root, because understanding is already settled by the time coding begins.

There’s a valuable side effect: by answering the questions, you, too finds holes in your own reasoning. We often think we have a clear idea, but it only becomes clear when someone asks, “What about this case?” The grill-me forces you to think about the edge cases you skipped. That’s why it’s useful even when AI no will write the code in the end—just going through the interview improves your plan.

The grill-me destination has a name: shared understanding. It’s the state where you and the AI see the problem the same way—with no hidden assumptions waiting to blow up. The interview doesn’t end when the AI gets tired; it ends when understanding converges. That’s why the skill says "until we reach a shared understanding": the stopping criterion is alignment, not a fixed number of questions.

David (Pocock’s colleague in the transcript) has a widely used variation on the same principle: "describe my vision, list out the 10 most consequential decisions (software design / architectural / product) that will shape this project, and interview me until you understand 98%." — describe my vision, list the 10 most consequential decisions that will shape the project, and interview me until you understand 98%. Notice the engineering: forcing AI to list the 10 most important decisions ensures the interview focuses on what matters, not trivial details; and the "98%" gives you a concrete convergence goal. It’s in the same family as grill-me—adversarial, in dialogue, until alignment. You’ll see this complete prompt, ready to copy, in the module 5.2 of the Solutions track.

The grill-me is a starting point, not a dogma. Since it’s just a skill of text, you can save it in your project and adapt it. Remember the golden rule from topic 3: adjust sparingly, because adding too much dilutes it. A few adjustments that fit without ruining the effect: ask for a summary of understanding at the end (“when we’re done, write a paragraph about what we agreed on”); limit the scope (“focus only on architecture decisions”); or set a cap (“maximum of 8 questions”). But start by copying the exact text—build your intuition before changing anything.

It’s worth remembering the philosophy of Pocock that runs throughout the whole track: he prefers procedures and want to be in the driver’s seat— "I know my skills, I don't want to delegate my thinking" (I know my skills; I don’t want to delegate my thinking). grill-me fits this perfectly: it doesn’t take you out of the driver’s seat; it sharpens your command. And it's the first link in a chain you'll build in the next module: grill-me → PRD → issues. Instead of dumping everything at once, you use grill-me to align, turn that alignment into a PRD, and the PRD into tasks—with you always in the driver’s seat.