In my AI Adoption Journey Post I mentioned the /grill-me skill by Matt Pocock. Being a big fan of it in the beginning, I kind of entered a through of disillusionment more and more. But Why?

Overgrilling leads to Premature Abstraction

More and more I ended up in situations where the grilling sessions seemed to never stop before answering around ~80 questions. Coming from a much more laid-back style of prompting, this initially seemed to be exactly what my agent driving projects needed and definitely improved my prompting in general.

What I actually found quite distracting was the fact, that the skill tries to walk you down the whole design tree of a problem, while I was rather trying to build small MVPs for quick feedback and iteration.

In addition to that, the grill-me skill often added new feature ideas to the thing I tried to build, which I actually liked but also led to distraction on the way to MVP. Not to mention that te skill also tried to add unnecessary hardening, e.g. thinking about complex deployment scenarios right from the first prompt, or already introducing an abstraction, multi-tenancy, or make things re-usable.

So when I lately attended a Hackathon, I reached a point where I realized I need some leaner version of this, to not spend the whole Hackathon answering questions but rather ship something.

Meet /lean-grilling

Rather then focussing on all aspects walking down a design tree, the lean-grill skill will ask you only the questions needed to unblock the next iteration of an idea.

---
name: lean-grill
description: Run a lean grill for MVP plans, decisions, and ideas.
---

Run a lean grill: sharpen the next buildable version, not the perfect long-term architecture.

First, inspect available facts from the environment instead of asking for them.

Then ask one question at a time. Each question must target the next unresolved decision that blocks implementation or materially changes the MVP.

For every question, include:
- Your recommended answer
- Why that answer is good enough for the MVP
- What risk or future decision it parks

Prefer fast, reversible decisions. Do not walk every branch of the decision tree unless the branch affects the next one or two iterations.

If a concern is real but not MVP-blocking, add it to “Parked risks” and continue.

Stop when we have enough shared understanding to build the first useful version. Then summarize:
- MVP scope
- Decisions made
- Assumptions
- Parked risks
- Recommended next action

Do not act on the plan until I confirm the summary.

This significantly reduced the number of questions being asked and enabled exploration of ideas and feedbacking quickly in favor of planning down a full blown design before even knowing the directions you want to take.

In addition to that, I also created the /lean-grill-with-docs skill which basically just combines with the domain modeling skill of matt pocock. This is still quite useful if you’re in an environment which has all the files needed for Matt Pocock’s (e.g. CONTEXT.md ). ´

When to use OG /grill-me ?

I still use the original /grill and /grill-with-docs skills. Usually when what I try to build is a fundamental concept on an existing codebase or touches several modules, I rather take the time to fully grill the thing.

If I just want to try something real quick, some kind of a mini PoC of what I imagine, I definitely use the /lean-grill

In fact, you can always use the outcome of a lean grilling session as input of a mature grilling session, once you need to extend the feature.

Prototype Skill

When trying to describe an application UI, I find the /prototype skill quite useful in combination with lean-grilling. It will generate 3 (or more if you ask for it) versions of a user interface for your application, which you can quickly cycle through.

Once you decide vor a variant, this will implicitly answer questions you’d answer text based usually

My Skills Repo

With the introduction of the /lean-grill skill, I decided to create a repo for my personal skills. You can find them here: https://github.com/eimerreis/skills