Matt Pocock’s 4-sentence “grill-me” skill became the most viral Claude Code skill (151K views) because it fills a gap no other intervention covers: the DURING-planning phase between premise-checking (“should we build this?”) and plan-review (“does this plan hold up?”). The skill instructs the agent to “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.”
The mechanism is Socratic branch-walking: systematically traverse a decision tree through adversarial questioning, where the interviewer recommends but the human decides. This directly combats First conclusions become nearly permanent — the brain resists its own updates — by forcing explicit consideration of every branch, it prevents the Inconsistency-Avoidance Tendency from locking in whichever design the planner thought of first. The pattern applies Invert, always invert — many problems are best solved backward to planning: instead of asking “what should we build?”, it asks “what would break if we chose this branch?” at every node.
Branch-walking is convergent (narrows toward a decision) AND exhaustive (covers all branches), which distinguishes it from brainstorming (divergent, generates options) and review (post-hoc, validates a completed plan). It produces the same benefit as Verification is the single highest-leverage practice for agent-assisted coding but applied upstream — verifying the plan before code exists is cheaper than verifying the code after. The resolved decision tree then feeds directly into Autonomous coding loops need small stories and fast feedback to work, where each resolved branch becomes a small, well-scoped story with clear acceptance criteria. Pocock uses it “even outside of coding,” suggesting the pattern is domain-general: any complex decision with branching dependencies benefits from adversarial traversal before commitment.
Pocock’s own diagnosis of why this skill was necessary, given later at an AI Engineer conference talk, grounds it in Frederick P. Brooks’ The Design of Design: when more than one party designs something together, there is an ephemeral shared idea floating between them — “not an asset, not something you can put in a markdown file” — that Brooks calls the design concept. The actual failure behind “the AI did something totally different than what I wanted” isn’t a prompting problem; it’s that “me and the AI don’t share a design concept” yet. Grill-me’s relentless interview is the mechanism for reaching that shared concept BEFORE any planning artifact exists — which is why he considers it superior to a coding agent’s default plan mode: “plan mode is extremely eager to create an asset, it really wants to just create a plan and start working, whereas it’s a lot nicer to reach a shared design concept first.” This reframes the skill’s real target as the pre-artifact stage, not the artifact’s quality — a distinction default plan-mode workflows collapse by rushing straight to the document. The resolved shared concept then feeds a companion practice, A ubiquitous-language file cuts AI verbosity and keeps implementation aligned with the plan — a standing glossary kept open during the same grilling sessions so the vocabulary, not just the decisions, stays shared.
Note (2026-07-09): The final paragraph above was added from a second source — Matt Pocock, “Software Fundamentals Matter More Than Ever” (AI Engineer conference talk) — which explains the design-concept rationale behind the grill-me skill described in the original entry. See sources/mattpocock-software-fundamentals-matter-more-than-ever.md.
A third source names why branch-walking’s exhaustiveness has become load-bearing rather than optional: The map is not the territory — with a capable-enough agent, finding your unknowns becomes the real bottleneck (Thariq Shihipar) — a more capable, more autonomous agent covers more territory per session, so the same plan now exercises far more unspecified decision points than it used to. Branch-walking resolves the unknowns you already know are open; the map/territory framing is what tells you an unknown exists at all before the agent finds it for you the hard way.