For a long time, the default move with any non-trivial feature was Claude Code's plan mode โ sketch the idea, let it draft a roadmap, skim it, hit execute. It felt productive. The output looked structured and confident. But as one developer documented in a widely-shared writeup this week, that confidence often masked a quiet inheritance of every gap in your own thinking. The real problem wasn't execution. Plan mode assumes your idea is already sound โ that the only variable is how well it gets built. Most of the time, that's the wrong assumption to start from. You can burn three days producing a clean, well-structured mistake, and plan mode will happily help you do it. That's exactly what makes /grill-me different. Developed by Matt Pocock โ the TypeScript educator behind Total TypeScript โ this skill generates zero code. It just won't let you start until you've actually thought through what you're building. The entire prompt is three sentences: "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." No clever machinery โ the value lives entirely in being forced to answer.
A Grill Session in Action
Pocock opened his personal .claude directory as a public skills repo on February 3, 2026. It crossed 50,000 GitHub stars within three months, peaking near the top of GitHub's trending charts โ not because it's a framework or wrapper, but because it solves a universal failure mode in AI-assisted development: misalignment that only surfaces after implementation. The workflow runs as a continuous pipeline โ /grill-me โ /to-prd โ /to-issues โ /afk. Each step feeds the next, and the rule is simple: never /clear between stages. The PRD skill leans on everything from the grill session โ your answers, your reasoning, the edge cases you named, the trade-offs you accepted. A PRD generated cold is a template. One generated right after a proper grill has a point of view.
Beyond the Grill
The repo includes several other skills worth adding to any Claude Code stack. /handoff compresses an active session into a document a fresh agent can pick up โ protecting deep grill context when you need to prototype something without derailing the interrogation. /tdd builds vertical slices one failing test at a time, turning acceptance criteria from the grill into tests Claude has to satisfy before claiming anything works. /zoom-out pulls context up a level for unfamiliar code stretches, and /grill-with-docs tests your plan against an existing domain model โ the right call when extending live systems. Pocock's writeup also covers infrastructure habits that keep Claude Code lean around the grill workflow: run grill sessions on frontier models like Opus where question quality matters most, then switch to cheaper models like Sonnet for implementation. Use .claudeignore (same syntax as .gitignore) to cut context bills โ excluding .next/ alone can trim 30โ40% of context in a Next.js repo. Keep files under 200โ300 lines, because agents read whole files and a large file is a tax every agent pays, every time it looks.
The Leverage Has Moved Upstream
The shape of building software has flipped. Writing code used to be the bottleneck โ now it's the part AI agents handle best. What humans bring to the table is deciding what to build in the first place and being precise enough that implementation doesn't drift. /grill-me earns its stars by enforcing exactly that discipline, at machine speed.
Key Takeaways
- Plan mode assumes your idea is sound; /grill-me tests whether it is before any code exists - The full workflow chains grill โ PRD โ issues โ execution without clearing context between steps - /handoff protects deep sessions from tangents โ fire side quests to fresh agents, return to the grill clean - Run grill on frontier models for sharper questions; implementation on cheaper ones to control costs .claudeignore and file size discipline compound across every read in every session
The Bottom Line
The highest-leverage skill in this workflow generates no code at all โ because specificity and edge-case honesty are now the scarce resource, not output. If you're still using Claude Code as a yes-man, you're paying for an expensive autocomplete. /grill-me proves the real value is in finding out what you didn't know you didn't know.