Pressure-Test Agent Decisions with grill-me

Where this fits

You've drafted a migration plan, picked a database, or settled on an API contract. Before you commit, you want an agent to poke holes in it—but generic "review this" prompts often produce vague praise or a list of nitpicks. grill-me steps in when you need a focused adversarial pass: it identifies the single most consequential assumption, asks one pointed question at a time, and ends with a clear recommendation and the evidence that would change it.

Why agents benefit

• Zeroes in on the highest-risk assumption rather than spreading attention across every minor detail, so the conversation stays tractable.
• Asks one question at a time, forcing a Socratic dialogue that surfaces hidden trade-offs without overwhelming the user.
• Challenges vague claims with concrete counterexamples or failure modes, making the pressure test actionable instead of philosophical.
• Reads code or docs first to avoid asking for facts it can establish itself, saving user time and grounding the critique.
• Ends with a crisp unresolved-decision summary plus the evidence that would flip the recommendation, leaving a clear next step.

Practical scenarios

Validating a database choice

You're torn between Postgres and DynamoDB for a new service. grill-me reads your schema sketch and traffic estimates, then drills into the write-hot partition key assumption, forcing you to justify the access pattern before recommending a fallback hybrid.

Stress-testing a refactor plan

You've outlined a three-phase refactor to decouple a monolith. grill-me inspects the interface boundaries and asks whether the new abstraction leaks implementation details—then gives a concrete failure mode and a safer alternative.

Auditing a rollout strategy

Before a canary deployment, you want to know what breaks if the new version mishandles a rare edge case. grill-me walks through the failure path, identifies the single untested assumption, and suggests a rollback trigger you hadn't considered.

Add it to your agent workflow

Install the skill with:

npx skills add jyt6640/persona-harness --skill grill-me

Then invoke it explicitly when you need a decision review:

/grill-me Should we move our batch job to AWS Lambda?

The agent will inspect your current code, ask one focused question, and give a recommendation with trade-offs—without creating tickets, plans, or branches.

Compare before adopting

Before you rely on grill-me, compare it against alternatives: deep-interview for ambiguous product outcomes, ralplan for adversarial review of approved high-risk plans, and the standard review or debug skills for code-specific checks. Check the repo's freshness (last commit, issue activity) and whether the maintainer responds to feedback. Also verify it fits your agent's permission model—grill-me is conversation-only and mutates nothing, which is a plus if you want zero side effects.

Why it is worth tracking

At 20 stars, grill-me is early but shows discipline: a narrow scope, explicit trigger rules, and a clear handoff to other skills. That focus suggests the author understands agent behavior beyond demos. Watch for community adoption and whether it evolves to handle multi-question sessions or integrates with your existing tooling. For a quick evaluation, the skill page at OpenAgentSkill includes usage examples and maintainer history—worth a look before you wire it into your daily workflow.