I was debugging a pipeline. I had found the problem. It was one line in a config file. I typed my prompt fast: "Change the line where the timeout is set to use the environment variable instead of the hardcoded value."
What happened next cost me 80,000 tokens.
The model did not change one line. It searched the whole codebase. It read file after file. It tried to figure out which timeout I meant. Was it the HTTP client timeout? The database pool timeout? The retry timeout in the queue processor? There were eleven places in the project where a timeout was set. The model checked every one. It looked for hardcoded values. It compared them to environment variable patterns. It wrote edits and threw them away.
It did exactly what I asked. I had not pointed at a line. I had not named a file. I had not said which timeout. I thought the model would know what I meant. I was wrong.
That mistake cost me 80,000 tokens. At Haiku pricing that is about $0.24. But the real cost was bigger. It broke the way I thought LLMs work. I had been treating the model like a coworker. I thought it shared context with me. I thought it could read between the lines. I thought it would guess the right timeout from the conversation. The model is not a coworker.
The model is a Genie.
I mean this for real, not just as a story. In the Aladdin tale, the Genie grants wishes perfectly. It does exactly what you say. It does not care what you meant. You wish for a million dollars. The cash appears. But it is locked in a vault you cannot open. Or it is taken from someone you love. Or it is fake. The Genie heard every word. The Genie missed every meaning behind the words.
LLMs work the same way. They run prompts with great skill. They do not guess what you really mean beyond what you wrote. When I typed "change the line where the timeout is set," the model heard: "there is a line with a timeout; I can change it; I should find it." It did not hear what I left out. The file name. The variable name. The exact change I wanted.
I had been making wishes the way people do in fairy tales. Loose wishes. Hoping for a kind reading. Trusting the Genie to want the best for me. But the Genie does not care about your wellbeing. The Genie just runs the wish.
After the incident, I started listing my bad wishes. Every vague prompt I had written was a monkey's paw. That is a wish that comes true in words but misses the point completely.
"Refactor this function." What does refactor mean? Rename variables? Pull out helper functions? Flatten the nesting? Change the signature? The model picks one option. Or it picks all of them. Or it asks questions that burn tokens. I never said what I wanted.
"Make the output cleaner." Cleaner how? Shorter? Better formatted? Fewer warnings? The model guesses. Sometimes it guesses right. Often it does not. Then I spend more tokens fixing the fix.
"Write a blog post about X." What length? What audience? What angle? What tone? What am I not allowed to say? What must be included? The model writes something generic. Generic is the safest choice when the wish is vague.
Every one of those is a monkey's paw wish. I wrote them. Then I felt frustrated when the output was wrong. But the frustration was misplaced. The Genie did its job perfectly. I just made a terrible wish.
Precise wishes are different. They give the Genie everything it needs to do what you really want. They state the context. They state the constraints. They state the format. They state who is answering and what the task is in plain terms. They leave nothing for the Genie to guess.
I started thinking about what a precise wish needs. Every time I looked at a bad prompt, the same things were missing. Something about who was speaking. Something about the surrounding context. Something about what was not allowed. Something about what the output should look like.
I turned this into the sinc format. It has six bands. Together, the six bands fully specify a prompt. The idea comes from the Nyquist theorem, which says you need at least 2x the signal frequency to rebuild it cleanly. If you leave parts of a prompt out, the model fills those parts in by guessing. That guessing is called specification aliasing.
The six bands are PERSONA, CONTEXT, DATA, CONSTRAINTS, FORMAT, and TASK. Each one holds a share of output quality. CONSTRAINTS holds 42.7%. That is the biggest share by far. It is not a surprise. CONSTRAINTS is where you stop hoping the model will figure out your limits. You write them down instead. You say what must not happen. You say what length is needed. You say which assumptions are wrong. You say what the model will want to do that you do not want it to do.
A Genie with no constraints is dangerous. A Genie with clear, complete constraints becomes a precise tool.
The formula above is not just decoration. It shows why a missing band cannot be made up for by a stronger band somewhere else. The bands multiply each other. If CONSTRAINTS is empty, the whole product goes to zero. It does not matter how good your TASK description is. The Genie hears a wish with a hole in it. It fills the hole however it wants.
I stopped treating prompts as conversation starters. I started treating them like contracts. The same way I write a function signature. The prompt is the whole interface. There is no room for things left unsaid.
I built the sinc format into sincllm.com as AI Transform. It takes a raw prompt and breaks it into the six bands automatically. It uses a locally fine-tuned Qwen2.5-7B model running at 290 tokens per second on my RTX 5090. Zero API cost. The model does the same breakdown I would do by hand, in under a second, every time.
Across 275 observations, I found this: structured prompts cut token cost by 97%. They also improve output quality so much that the first response is almost always the final response. The Genie stops searching because the wish is precise. The Genie runs the wish and stops.
I still think about that 80,000-token incident. Not with frustration. With gratitude. It was the most expensive lesson I have paid for in this space. It was worth every token. It made me stop expecting the model to think like a person. It made me start writing wishes the way a Genie needs them: complete, clear, and exact.
The Genie will not read between your lines. Write the lines.
// Production AI Engineering
sinc-LLM designs, audits, and stabilises production AI infrastructure: from vendor evaluation and cost accountability to incident controls and MCP architecture.
See what we do →