🤖 The Model Is the Same for Everyone
Start with the uncomfortable realization: if you use Claude Opus 4.8 and your competitor does too, you’re starting from exactly the same place. The same prompt, on the same model, tends to produce the same kind of response. The model has become commodity — identical input for everyone.
New here? One model (or LLM) is the AI that runs behind things like ChatGPT and Claude. "Same model" means the underlying intelligence is literally the same for everyone who uses it.
💡 The central idea
If the model is the same, it can't be your differentiator. Chasing “the perfect prompt” means optimizing the part everyone already has in common. The differentiator is somewhere else — and the next topics show you where.
The model is the same input for everyone.
Same prompt → same class of output.
Refining the prompt alone yields little.
The advantage is in the context.
🎨 Context, taste, and voice
What makes a result your is what you add to the model: your context, your taste (taste), your voice, and your decisions. That’s what makes the output stop sounding like "any AI" and start sounding like you or your company.
What is “context”? It’s all the knowledge you give the model beyond the request itself: how you make decisions, examples in your style, and rules for your business. In Claude Code, this usually lives in skills e docs. A skill is a reusable piece of instruction that the model loads when the topic comes up.
✗ Without your context
- ✗Correct text, but generic—“sounds like AI.”
- ✗Decisions that reflect the world's average, not yours.
- ✗You rewrite everything at the end to make it sound like you.
✓ With your context
- ✓Output already comes out in your voice and according to your rules.
- ✓Decisions reflect how you work.
- ✓Less rework: approve more, rewrite less.
What you provide beyond the request.
Your taste and quality standards.
The way it sounds like you.
Your choices, not the average.
🧠 The Problem Is Extraction
If context is the advantage, why isn't it in your system? Because the truly difficult part is extraction: get knowledge out of your head, where it lives in tacit and disorganized, and put it into a format the AI can use.
Tacit vs. explicit knowledge. Tacit is what you know how to do but can’t explain right away (riding a bike, "sensing" that a price is wrong). Explicit is what’s already written down and organized. grill-me exists to turn tacit knowledge into explicit knowledge.
🔍 Why extraction hurts
- • You already know so much that skips steps without realizing it — the “curse of knowledge.”
- • A lot only comes up when someone asks directly about it.
- • Explaining your own process reveals parts that you don’t fully understand yourself as well as you thought.
You know it, but you haven’t written it down.
Already organized and readable.
Extraction, not the model.
Those who know skip steps.
💨 A brain dump isn't enough
The natural reaction is: “I’ll dump everything into Claude for 5 minutes and that’ll be good enough.” It’s never good enough. Free-form dumping is held hostage by your own blind spots—you talk about what you remember, generalize nuances, and don’t even notice the gaps.
✗ 5-minute brain dump
- ✗You decide what’s relevant — and leave out the rest.
- ✗It generalizes the nuance (“it depends”) without explaining further.
- ✗Gaps stay invisible: no one points them out.
✓ Guided interview
- ✓The questions draw out what you wouldn't bring up on your own.
- ✓Each branch of the decision is followed through to the end.
- ✓Gaps become questions—and then flags.
🎯 Tip
When you catch yourself thinking "I’ve explained enough," that’s exactly when grill-me becomes valuable: the next question often reveals something you would never have written spontaneously.
You don’t know what you left out.
Generalizing kills the detail.
The question exposes the gap.
It’s the sign that you’re stopping too soon.
🪓 Sharpen the Axe
The quote attributed to Lincoln sums up the bet behind grill-me: "If I had 6 hours to chop down a tree, I’d spend the first 4 sharpening the axe." Grilling is the sharpening stage—investing time up front so everything afterward goes much more smoothly.
The 4 hours sharpening (it seems unproductive)
A grill-me session is slow, repetitive, and sometimes uncomfortable. You get the feeling that “this is a waste of time.”
The 2 hours cutting (the real work)
With the extracted context, the skill or project comes out much more right on the first try. The cut is quick because the axe was sharp.
⏳ Productive patience
The cost of grill-me is felt now; the return appears afterward, spread across every time you use the skill or the resulting doc. It’s an investment, not an expense.
Tedious, but brief.
With each future use.
It’s not an expense; it’s capital.
Preparation beats brute force.
📈 The iteration curve
No skill is perfect from the start. You iterate. The question is what level you’re starting from. Without grill-me, iteration 1 starts at around 70% accuracy and improves slowly. With grill-me, you jump to around 90% in the very first version—and you still iterate, just from a much higher starting point.
Look at the two starting points on the left: the time spent grilling points to gains that would otherwise take many iterations. The ceiling (~95%) never reaches 100%—the skill evolves with you and the business.
70% vs. 90% in iteration 1.
Gets to “good” faster.
Never 100% — and that’s okay.
The skill changes with you.
📄 The capture file
Here's the heart of the practical version of grill-me: the file is the source of truth, not the model’s memory. Every answer is saved in a markdown file inside brainstorms/ — before the next question. The file, not the conversation context, keeps everything.
Context window? It’s the model’s “short-term memory” in the current conversation. It has a limit: in a long interview, the beginning may be summarized or forgotten. That’s why the checkpoint (writing to disk after every answer) is what keeps anything from getting lost.
# Packaging: Brainstorm / Discovery Notes Date: 2026-06-22 · Goal: extrair o processo de embalagem ## Summary / key decisions (síntese contínua, atualizada a cada resposta) ## Q&A log ### Q1 — fluxo de empacotamento - Asked: como você decide o tamanho da caixa? - Captured: regra "menor caixa que cabe + 2cm"; nas palavras dele - Flags: tabela de custo por caixa -> pedir ao financeiro ## Open flags (pending input) - tabela de custo por caixa -> financeiro
📌 Why disk beats memory
If the interview lasts an hour, the model goes conflate or forget details from the beginning as the context fills up. The file doesn’t forget. If the session crashes now, it already contains everything that’s been said so far.
A predictable home.
Save after every response.
File > context.
Lost your session? Nothing is lost.
🎯 When to use it (and when not to)
Grill-me is a tool for extraction, not a universal hammer. It shines when there’s dense, disorganized knowledge to draw out of someone. For a trivial, well-defined task, it just gets in the way.
✓ Use when
- ✓Want stress-test a plan or design before execution.
- ✓Is doing discovery with a client or stakeholder.
- ✓Want to extract a business process that lives only in your head.
- ✓Are you going to build or improve a skill and wants to start from the top.
✗ Avoid when
- ✗The task is obvious and already fully specified.
- ✗There’s no tacit knowledge to extract—just execution.
- ✗The answer is no codebase/doc (then you just read it).
- ✗You want speed on a mechanical step.
🧭 Rule of thumb
Ask: "Is there anything important that's only in my head?" If so, grill. If the answer is in a file or is trivial, skip the grill-me and get straight to work.
Find gaps earlier.
Extract from stakeholders.
Document what's tacit.
Obvious tasks don’t need one.
🧠 Module summary
Next track:
Track 2 · The Techniques — how the interview works under the hood: one question at a time, a recommended answer, a dependency tree, flags, and a checkpoint.