⌨️ How to invoke
There are two ways to trigger grill-me, and both work: the slash command (/grill-me) e a natural language. How the skill loads through your description, just describe what you want and it kicks in automatically.
Slash command? It’s the shortcut /nome-do-skill that invokes a skill directly in Claude Code. The alternative is simply to write the request in English — the skill is triggered by its description.
Objective: start a grill-me session on a topic of your own.
Grill me about. One question at a time, with your recommended answer, and checkpoint to a file after each response.
How to verify: it should create a file at brainstorms/, say where you saved it and do the first question — without dumping several at once.
The direct shortcut.
"Grill me about X."
It’s what triggers it.
Use the form that feels natural to you.
🏢 Grill the business/process
The most valuable use: extracting a your business process from beginning to end. It’s the kind of thing that lives only in the mind of the person doing the work — and when documented, becomes a reusable skill, training, or context.
Objective: map out an entire process, with flags for anything you don't know.
Quero que você me grile relentlessly sobre o processo de, do começo ao fim. Regras: - Uma pergunta por vez, com a sua resposta recomendada. - Resolva dependências em ordem (decisão de cima primeiro). - Se algo estiver num doc/código, leia em vez de perguntar. - O que eu não souber, registre como flag com o dono certo (ex.: " -> ") e siga. - Faça checkpoint em brainstorms/ a cada resposta. - No fim, releia tudo, reconcilie contradições e me dê um recap: capturado, flags em aberto, próximo passo.
How to verify: by the end, there must be a brainstorms/AAAA-MM-DD-embalagem.md with Summary, Q&A log, and Open flags — and the recap must list at least one flag with an owner.
🎯 Tip
Always mark the variable parts with <colchetes> before pasting. The more specific the scope (“packaging process", not "my business"), the question tree is better.
A process, not “everything.”
Serial, flags, checkpoint.
Marked in <brackets>.
Captured · flags · next.
😰 The difficulty of answering
Here's the little-spoken truth: being grilled is uncomfortable. The questions touch on exactly what you've never put into words, and answering is tiring. That friction isn't a flaw in the process—it's the process working.
💡 Why discomfort is the work
The question that makes you hesitate is the one that matters most: it found something you knew how to do but had never explained. If nothing felt uncomfortable, grill-me wouldn’t be extracting anything new—it would just be confirming the obvious.
✗ The reaction that undermines you
- ✗"That’s enough, it’s good enough" at the first sign of fatigue.
- ✗Give a shallow answer just to get past the question.
- ✗Treating the hard question as a nuisance instead of gold.
✓ The approach that extracts
- ✓Treat discomfort as a sign that you've found something.
- ✓Let the session run — it could be 5 questions or 30.
- ✓Answer with the real nuance, even if it takes longer.
⏳ Set expectations
A good session can take more than an hour. That’s normal and good. The sign you should keep going is precisely the feeling of “I’ve already explained this” — there’s almost always nuance missing there.
The hard question is gold.
"Enough" is deceptive.
5 to 30+ questions.
Worth the wait.
🤷 When You Don’t Know
At some point, grill-me will hit a point where you don’t explain as well as the operator who does that every day. It’s not a failure — it’s information. It becomes a flag for the right stakeholder, and you get the answer later.
Objective: turn “I don’t know” into an actionable flag without blocking the session.
I don’t know this step in detail—the person who knows it best is the. Record it as a flag with that owner and move on to the next question. I'll bring her response afterward, and we'll update this brainstorm.
How to verify: the section Open flags the capture file gets an item <passo do estoque> → <operador> and the interview continues instead of getting stuck.
🔁 The “I don’t know” cycle
- 1. The question touches on a gap in your knowledge.
- 2. It becomes a flag for the person who knows (operator, finance, legal…).
- 3. You find the person and collect the answer.
- 4. Re-grill the brainstorm with the new information (next topic).
Show where you’re not an expert.
Operator/stakeholder.
It brings the answer later.
Don’t make up what you don’t know.
🔄 Grill again
How each session leaves a document, you can return to it when you discover something new. The brainstorm is alive: found a process improvement, solved a flag, changed your mind? Grill again and update it.
Objective: update an old brainstorm with new information and resolve flags.
Leia brainstorms/<2026-06-22-embalagem.md>. Me grila de novo só sobre o que mudou:e a flag " " foi resolvida — o valor é . Atualize o arquivo (corrija as entradas afetadas) e me diga o que ficou diferente.
How to verify: the same file is updated (not a new one), the flag comes from Open flags, and the Summary reflects the change.
📌 Living doc, not frozen
In the reference video, that’s exactly what happens with the packaging brainstorm: when you find a “major breakthrough,” you go back to the doc and say, “here’s what I found, let’s update all of this.”
Come back whenever you like.
Grill on the delta, not everything.
Closes what was open.
Update, don't duplicate.
📦 From brainstorm to deliverable
Capture is raw material, not the final product. The raw always stays in brainstorms/; when the session produces a polished deliverable—a plan, a map, a spec, a skill—that artifact can move to the corresponding project folder.
Raw stays in brainstorms/
The raw interview log—Q&A, flags, synthesis—lives in a predictable place and never gets scattered.
Polished becomes a deliverable
From the raw material, you distill a clean plan/spec/skill—and that one does go in the folder projects/ of the work.
The return path remains
If someone asks "where did this decision come from?", the original brainstorm is still there and traceable.
Always in brainstorms/.
Move to projects/.
Raw ≠ deliverable.
The origin is recorded.
⚙️ Grill-me + skill-creator
This is where the cycle closes. At the end of a session, grill-me realizes you have a skill and a doc with nuance that isn’t in them — and offers to update both. Combined with a skill-creator, is what takes a skill from ~70% to ~95%.
The nuance grill-me extracts doesn’t stay in the brainstorm: it feeds into updates to the skill and the doc. That’s “sharpening the axe” turning into a real cut.
🔗 The closing offer
In the end, grill-me says something like: "I noticed you have this guide and this skill, and there’s a lot of nuance here that isn’t in them — do you want me to update both?". You answer "yes"—and your OS gets smarter.
Brainstorm → skill+doc.
"Do you want me to update it?"
Refines the result.
The leap in quality.
⛔ Anti-patterns
To wrap up, the mistakes that drain a grill-me session. Each one destroys one of the safeguards you learned in the previous tracks.
✗ Anti-patterns
- ✗Answer in batches — kills the serial rhythm.
- ✗Let skip the checkpoint — risks losing everything.
- ✗Ask what you can explore no code/doc.
- ✗Stopping too early at the first "it's good enough."
✓ The right path
- ✓One question, one answer, one checkpoint.
- ✓Require a checkpoint after every response in the prompt.
- ✓Let the grill-me read before asking.
- ✓Follow through until every branch is covered—flags resolve the rest.
⚠️ The worst of the four
Skipping the checkpoint is the most dangerous thing: in a long session, if the context fills up or the session crashes, everything that hasn’t been saved some. That's why the rule is non-negotiable—explicitly ask for the checkpoint in your prompt.
Always serial.
Save every response.
Before asking.
Cover every branch.
🎯 Module summary
You completed the course 🎉
Now you know why context is everything, how the interview works, what's inside the SKILL.md, and how to use grill-me for real. Next step: run the prompt from section 2 in one of your own processes.