AnthropicWorkflows

· Beginner

Illustration: a round charcoal kettle grill with its lid closed, standing on three legs with two wheels, seen straight-on, in black and white.
Illustration

Reported, not yet confirmed.

Let a skill grill you before you build anything

An investor runs Matt Pocock's free grill-me skill almost daily so Claude asks hard questions before building.

You can make your agent interview you before it builds, so the gaps show up before the work does. @atomicalcapital wrote on Oct 22: "A skill I use almost every day: grill-me from Matt Pocock. It's simple and genius." On their latest run: "In total, Claude asked me 10 questions. More than he typically would without the skill." This is self-reported. We have not run it.

Disclosure: @atomicalcapital's bio says "Ex-nuclear engineer hunting hidden atoms. Teaching investors to wield Claude as a junior analyst." The skill is free and MIT-licensed on GitHub; its docs link to pages on aihero.dev.

What it replaces#

Before: the agent starts building on your first description and you find the open questions late. After: a short interview first. The post says it works "When you build something in Cowork or Claude Code".

How it works#

  1. Install it. For Claude Code, the README gives claude plugin install mattpocock-skills@claude-plugins-official, which installs the whole set. Copying files by hand? grill-me is a one-line file that hands off to a second skill named "grilling", so copy the grilling skill too. That second file holds the actual interview.
  2. Call it yourself. "You invoke this by typing /grill-me; the agent won't reach for it on its own," per the docs.
  3. Start fresh. The docs say: "Start it in a fresh conversation, not on top of a plan you already had an agent write."
  4. Answer in rounds. The grilling skill asks every question whose prerequisites are settled at once, and tells the agent to "number each question and give your recommended answer." Each question is worded so that answering yes accepts that recommendation.
  5. Push back. The docs warn that "The failure mode is passivity". Say "I don't know" when you don't.

Good to know#

  • Expect more than 10. "Forty-six questions across four rounds is an ordinary session," per the docs.
  • The docs say to leave plan mode off, because "Plan mode makes the agent hurry to produce a plan".
  • If it asks hundreds of questions, the docs say "Usually the scope was too large." Split the work first.
  • It "needs no repo" and the subject "doesn't have to be code," so it works for a memo or a business call.

Same idea in Grok Bot#

Our read: Grok Bots can keep skills (the changelog mentions "Bots with a lot of saved memory or skills"), so a Bot can follow the same instructions. Paste the grilling file into chat, ask your Bot to save it as a skill, then ask for it by name; per the changelog, "Typing / in the composer no longer opens a menu of skills and actions." Or use this prompt (ours, untested):

Prompt
Before you do anything, interview me about this idea until we
both could defend every choice. Ask in rounds: each round, ask
every question whose prerequisites I've already settled,
numbered, each with your recommended answer. Skip anything that
depends on an answer you don't have yet. Don't write a plan or
start work until I say "go". If I say "I don't know", note it
as an open question.

Sources#

Sources

  1. x.com
  2. github.com
  3. github.com
  4. github.com
  5. github.com

Something wrong? Tell the bot