Lately I keep seeing posts from developers who spend hundreds of euros a month on LLM tokens. They all ask the same thing: is it worth it?

I'll admit it. A few months ago, I was doing the exact same thing. My token bills kept going up. I blamed the models for raising prices. The real problem was my workflow.

The problem is not the model

Let's be real. If you're throwing hundreds of euros at LLM costs, you probably need to check your prompts first.

Stop using the most expensive model to change a single line of CSS.

A pricey model does not fix a bad prompt. It just makes a bad prompt cost more.

Two halves of the bill

Token waste splits into two places.

Input tokens: you keep re-explaining your project because the model never learned your language.

Output tokens: the model opens with "Sure! Happy to help" and wraps one useful line in five filler sentences. You pay for all of it.

Fix one and you save a bit. Fix both and it starts to add up. That is the whole story behind my 60% drop.

What I changed

I combined /grill-with-docs and /caveman and together they cut my tokes by about 60%. But not only that it improved my speed and the output got better too.

Here is how each one works.

/grill-with-docs is built by Matt Pocock, an educator, content creator and engineer. Most token waste does not come from one bad answer. It comes from the back and forth. You ask something, the model guesses what you mean. You correct it, it guesses again.

Each round costs input and output tokens. Most of that cost is the model re-learning your project from scratch.

/grill-with-docs fixes that at the source. It reads your repo first, finds the words that actually matter, and asks you short, pointed questions until everything is clear. Then it writes that down in a file called CONTEXT.md.

The workflow has four steps: orient, mine, grill, write. Orient finds where context already lives, or creates the first CONTEXT.md. Mine searches the codebase for domain terms: type names, API fields, event names, user-facing copy. It looks for places where the same word means two different things.

Grill is the important part. The model asks one question at a time, tied to real evidence from your code. Not "what is an order?" but "in checkout.js you call this Purchase, and in billing you call it Order. Which word should we use everywhere?"

That sounds slow, but it is cheap compared to fixing wrong assumptions later. One clear answer now can save ten correction later.

Why they work better together

/grill-with-docs reduces the rounds of back and forth and makes sure the model knows your project inside out. /caveman reduces word count and makes the model answer shorter and more to the point. One fixes input quality over time and the other fixes output size on every reply.

Say you already have CONTEXT.md defining your terms. Caveman answers get even shorter because the model already knows what you mean.

What you can learn from this

You do not need these exact skills. You need the ideas behind them.

Write down your project overview once, before you ask the model to build features. Ask one precise feature/change at a time. Pick canonical names and stick to them.

Then tell the model to answer short. No "Sure! Happy to help." opener. My girlfriend would say, "React like a Dutchman" directly to the point. The LLM can do the same.

Match the tool to the task. A CSS tweak does not need your most expensive model. Put time into creating a better workflow, this will save you time, frustration and money in the long run.

Better prompts beat bigger models, every single time!