← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Conceptupdated Sat Sep 26 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Grill Me

promptingskillsdesignelicitationspec-driven-developmentpocockinterview-pattern

Grill Me

Matt Pocock's viral skill (~13,000 stars) and the vault's name for a broader pattern: make the model interview you relentlessly before it writes anything.

The instruction shape: interview me relentlessly, walk down each branch of the design tree, resolve dependencies one by one — then build.

The problem it solves

Pocock's diagnosis (Software Fundamentals Matter More Than Ever (Matt Pocock, AI Engineer)) is failure mode #1 of AI-assisted building: "AI didn't build what you wanted."

His framing borrows Brooks: what you actually need to transfer is the design concept — the ephemeral idea you share with a collaborator. It lives in your head, it is not in the prompt, and a model that starts generating immediately is building from a partial map. Grill Me forces the map to be shared before code exists.

The independent convergence

The pattern's strength is that three unconnected voices arrived at it:

Voice Form Source
Matt Pocock Grill Me — interview me, walk the design tree Software Fundamentals Matter More Than Ever (Matt Pocock, AI Engineer)
Thariq Shihipar (Anthropic) "ask the model to interview me" — notes Claude Code's plan mode is "extremely eager to create an asset", prefers an interview-first flow Field Guide to Fable (Thariq Shihipar, Anthropic)
Nate Herk the same interview idiom in an automation register vault's Herk thread

warning Contradicts the "independent convergence" framing: in AI Skills with Matt Pocock (The Pragmatic Engineer) Pocock says he got the idea from Shihipar ("a Tariq who works with Claude Code… get the agent to interview you") and encoded it as a skill. So Pocock ← Shihipar is a lineage, not independent discovery; only Herk (and the Economist curveball-question evidence) remain genuinely separate vantages. This weakens the triangulation argument above.

Shihipar's observation is the sharpest supporting evidence: the default behaviour of a strong coding harness is eagerness to produce, which is precisely what Grill Me suppresses. The skill exists because the tool's default is wrong for ill-specified work.

The empirical anchor

The Secret to Good Questions (Economist) supplies unexpected outside support: LLM-classified "curveball" analyst questions predict share-price movements, and systematised questions reduce interview bias. The vault reads Grill Me as the AI-native form of curveball-question elicitation — the value sits in questions that are not the obvious next one.

That connection also extends Higher-Order Thinking with an empirical hook: good questions are measurably informative, not merely good practice.

Where it sits in the SDD debate

Grill Me is Pocock's constructive half. He is the vault's main critic of naive spec-driven development ("vibe coding by another name"), but his alternative is not "no specs" — it is specs earned through interrogation inside a single session, at the codebase-quality layer, alongside Deep Modules and Ubiquitous Language.

Placed against the frameworks on Spec-Driven Development (SDD): BMAD and SpecKit supply planning and governance around the session; Grill Me supplies elicitation discipline inside it. They are complements, not rivals — which is why the Pocock/Dwarampudi disagreement is narrower than it first appears.

2026-09-26 — Field reports and when not to grill (AI Skills with Matt Pocock (The Pragmatic Engineer))

  • What it feels like: Gergely Orosz got 35 questions for a "simple" subscriber-check API endpoint — auth (bearer token vs JSON body), scope, exact rate-limit semantics — "intense and annoying and it forced me to think more." Pocock: "everyone's got a grill me story"; the skill is tiny but shows emergent behaviour — the model starts thinking outside the box and throwing ideas at you, like a senior engineer in the room.
  • What it's really for: closing the communication gap — "however smart the model is… it can't read your mind." Grilling establishes scope, non-scope, and your hierarchy of values, not just implementation detail. The human still makes every decision (vs /go-style prompts where the agent silently decides) and learns along the way — Orosz: "as long as we're learning I think we're fine."
  • Decision rule — shift right where you can:
    • Small enough to align after seeing it → don't grill.
    • Fits in one session but hard to row back from (wrong code in context poisons what follows) → Grill Me.
    • Spans multiple sessions → Wayfinder (Skill).
  • Downstream: grilling output → spec (destination document) → tickets (one per session) → AFK implement loop (Smart Zone and Dumb Zone, Day Shift Night Shift (Agents)). Variant grill-with-docs builds a Ubiquitous Language glossary during the grilling.
  • Waterfall objection: Pocock pairs upfront grilling with aggressive, cheap prototyping (build 3–4 versions, pick one) — which is what makes it not waterfall. Grady Booch (via Orosz): waterfall's sin was multi-year planning, not planning.
  • Scale: the parent mattpocock/skills repo is now ~230,000 stars (vault's earlier ~13,000 figure was Grill Me in mid-2026).

The one-line

Before a model builds, make it prove it has your map — the cheapest available fix for the most common AI-building failure.

Cross-references

Sources