Engineering disciplineClaude Code and agents that support the skills CLI

Matt Pocock / grill-me

What if the agent asked you the hard questions instead?

The plan you cannot defend out loud is the plan that will fail in week three.

What to remember

  1. Being questioned surfaces gaps faster than being advised.
  2. Answers you invent under pressure are the ones to double check afterwards.
  3. It pairs with brainstorming and writing-plans, which sit on either side of it.
  4. It cannot supply domain knowledge you do not have. It only finds where it is missing.

01What it changes in practice

Most plans fail on the parts nobody wanted to discuss: migration, error paths, who owns the data, what happens to existing customers. Those are exactly the parts a friendly review skips.

An interrogation surfaces them because it keeps pulling on branches rather than approving the shape. The discomfort is the mechanism, not a side effect.

02Where it disappoints

It can grind. Not every branch deserves resolution before implementation, and deferring some deliberately is a legitimate answer that you have to give explicitly.

It also cannot check facts about your organisation. It will accept a confident answer that happens to be wrong.

Install and first run

  1. Install it with the skills CLI from mattpocock/skills, published under MIT.
  2. Bring a real plan, not a topic. It interrogates decisions, not intentions.
  3. Write down every answer you had to invent on the spot. Those are the actual gaps.

Questions people actually ask

How is this different from asking for a review?
A review evaluates what you wrote. This questions you until the undecided parts of the plan are visible.
When should I use it?
Before implementation, on anything where a wrong decision is expensive to reverse.
Does it need a written plan?
Yes, bring a real plan. A topic produces vague questions and vague answers.

Sources

  1. mattpocock/skills on GitHubrepository
  2. Agent skills directory listingdirectory

Machine readable versions of this page