How the ‘grill‑me’ Skill Turns Vibe Coding Chaos into Clear Plans

When using Claude Code for "Vibe Coding", the AI tends to obey vague prompts blindly, guessing missing details and producing code that often requires extensive rework. The core issue is the AI’s over‑eagerness to act without clarifying the problem.

What is grill‑me ?

grill‑me

is a skill created by Matt Pocock (hosted in the mattpocock/skills repository) that flips this behavior. Instead of executing immediately, it asks the user a series of precise questions to surface hidden requirements before any code is generated.

The skill’s three commands, extracted from its front‑matter, are:

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, waiting for feedback on each
question before continuing. Asking multiple questions at once
is bewildering.

If a question can be answered by exploring the codebase,
explore the codebase instead.

These sentences make Claude act like an interview‑style reviewer, forcing the user to clarify each decision before proceeding.

Why “one question at a time” matters

The skill adopts the “design tree” concept originally described by Fred Brooks: every project is a hierarchy of decisions, each branching into smaller choices. By asking one question at a time, grill‑me ensures that the most upstream, ambiguous decisions are fixed first, preventing downstream rework caused by hidden assumptions.

Installing and invoking the skill

Installation is a single command: npx skills@latest add mattpocock/skills After the skill appears in the Claude Code skill list, you can start it with:

/grill‑me

Practical example: building an “auto‑hotspot assistant”

The author wanted a tool that automatically gathers daily AI‑related hotspots from platforms like X, Hacker News, and GitHub, formats them into a topic list, and pushes the list to Feishu. By feeding this high‑level intent to grill‑me, the skill first enumerated existing capabilities (agent‑reach, Feishu integration, scheduling) and then began a step‑by‑step interview.

During the interview, the skill asked ten+ focused questions, such as:

What AI sub‑category does the assistant target?

Which platforms should be scraped?

How should duplicate topics be handled across days?

One notable insight emerged when the skill warned that a daily run could repeatedly push the same hotspot if the assistant didn’t remember previous selections—a problem the author hadn’t considered.

After the interview, the author confirmed the plan, and grill‑me generated the necessary Claude Code snippets, set up Feishu webhook integration, and configured a scheduled task. The final result was a daily, curated list of AI‑related topics delivered to Feishu.

Comparison with Claude Code’s Plan Mode

Plan Mode tries to produce a plan quickly, filling in missing details on its own—essentially letting the AI “make decisions for you.” In contrast, grill‑me deliberately slows down, insisting on user confirmation for every decision, which reduces the risk of costly mis‑direction.

The recommended workflow is to run grill‑me first to clarify the problem, then hand the resulting plan to Plan Mode for rapid execution.

Limitations

grill‑me

is not a universal fix; using it for trivial tweaks (e.g., fixing a typo) can be tedious. It shines on tasks that require several hours of work and where a wrong direction would cause heavy rework.

Beyond code, the skill can be applied to product ideas, course outlines, or any scenario that involves a hidden decision tree, helping users surface considerations they might otherwise overlook.

In short, grill‑me transforms Claude from a “guess‑your‑mind” executor into a partner that forces you to articulate your thoughts clearly before any implementation begins.