Agent Skills: Grill Me

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when the user wants to stress-test a plan, get grilled on their design, or mentions \"grill me\".

user-invokedID: v1truv1us/ai-eng-system/grill-me

Install this agent skill to your local

pnpm dlx add-skill https://github.com/v1truv1us/ai-eng-system/tree/HEAD/skills/grill-me

Skill Files

Browse the full folder contents for grill-me.

Download Skill

Loading file tree…

skills/grill-me/SKILL.md

Skill Metadata

Name
grill-me
Description
"Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when the user wants to stress-test a plan, get grilled on their design, or mentions \"grill me\"."

Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.

Grill Me

Overview

Stress-test a plan, design, or decision through a relentless, one-question-at-a-time interview. The goal is not to win an argument; it is to walk every branch of the design tree until the user and the agent share the same mental model and the next step is obvious.

When to Use

  • The user says "grill me", "grill this", or "stress-test my plan"
  • A design feels half-baked and needs forced clarification
  • Dependencies between decisions are unclear
  • The team keeps revisiting the same questions
  • A spec or proposal needs validation before implementation

Trigger Phrases

  • "Grill me on this plan"
  • "Grill this design"
  • "Stress-test my approach"
  • "Interview me about the architecture"
  • "Poke holes in this proposal"

Execution Rules

  1. One question at a time. Do not dump a checklist. Ask a single focused question and wait for an answer.
  2. Provide a recommended answer. After asking, offer what you would recommend given the codebase and context. Frame it as "If I were making this call, I would..." so the user can accept, reject, or refine it.
  3. Walk the tree depth-first. Resolve the current branch before opening a new one. Track dependencies explicitly: "Before we pick the database, we need to decide the consistency model."
  4. Explore the codebase when applicable. If a question can be answered by reading code, tests, docs, or configuration, inspect the repo instead of asking. Cite file paths and evidence.
  5. Surface assumptions. Name what you are assuming before you ask the next question.
  6. Summarize before moving on. After resolving a branch, state the agreed decision and how it constrains remaining branches.

Suggested Branches

Use these as a default decision tree, adapted to the user's domain:

  • Goal and success criteria — What outcome must be true for this to be done?
  • Scope boundaries — What is in, out, and deferred?
  • Constraints — Time, budget, compliance, team, tech, and maintainability limits.
  • Dependencies — What must be decided first? What blocks what?
  • Trade-offs — What are we optimizing for? What are we willing to give up?
  • Risks and mitigations — What could invalidate the plan? How would we know?
  • Ownership and next step — Who decides, who implements, and what is the first action?

Example Invocation Flow

User: "Grill me on my plan to migrate the auth service to OAuth2."

Agent: "Before we talk about OAuth2 providers, what outcome must be true for this migration to be considered done? My recommendation: define a single acceptance test where a user can log in, revoke the session, and the old token format is rejected."

User: "Done means no more legacy sessions and all clients use the new flow."

Agent: "Got it. That means we need to decide how to handle active legacy sessions. I explored src/auth/session.ts and see we currently issue 30-day refresh tokens. I recommend a forced re-authentication at the OAuth2 cutover rather than backporting revocation. Do you accept that?"

User: "Accept."

Agent: "Next branch: which OAuth2 flow per client type? ..."

Verification

  • [ ] Every branch of the design tree has been resolved or explicitly parked
  • [ ] Each decision is recorded with the assumption it depends on
  • [ ] The user and agent can state the plan in the same words
  • [ ] The next action is concrete and owned

Anti-Rationalization Table

| Excuse | Counter | |---|---| | "I already thought this through" | Thinking privately is not the same as shared understanding. One forced question usually exposes a gap. | | "This is too detailed" | The detail is where plans fail. Better to surface it now than at 2 a.m. in production. | | "Can you just give me the answer?" | Answers without shared context get rejected or misapplied. The interview builds the context that makes the answer stick. | | "We can decide that later" | Later decisions often invalidate earlier ones. Park only when the dependency direction is clear. |