I was debugging a pipeline. I had found the problem — a single line in a config file that needed to change. I typed my prompt fast, the way you do when you're in flow and you know exactly what you want: "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 a line. It explored the entire codebase. It read file after file, following chains of imports, trying to figure out which timeout I meant — the HTTP client timeout? The database pool timeout? The retry timeout in the queue processor? There were eleven places in the project where timeout was set. The model worked through each one systematically, checking if it was hardcoded, cross-referencing with environment variable patterns it found elsewhere, writing and discarding candidate edits.
It was doing exactly what I asked. I had not pointed at a line. I had not named a file. I had not specified which timeout. I had assumed the model would know what I meant because I knew what I meant.
That assumption cost me 80,000 tokens — roughly $0.24 at Haiku pricing, but the real cost was deeper. It fractured my mental model of how LLMs work. I had been treating the model like a colleague — someone who shares context, who can read between the lines, who knows which timeout I'm probably talking about given the conversation so far. But the model is not a colleague.
The model is a Genie.
I mean this precisely, not as a metaphor. In the Aladdin story, the Genie grants wishes with perfect fidelity to what was asked and zero fidelity to what was intended. You wish for a million dollars and the cash materializes — in a vault you cannot open, or taken from someone you love, or counterfeited. The Genie heard every word. The Genie missed every implication.
LLMs work exactly this way. They execute prompts with extraordinary capability and zero inference about your underlying intent beyond what is written. When I typed "change the line where the timeout is set," the model heard: "there exists a line where a timeout is set; I have permission to change it; I should find it." It did not hear the part I left unsaid — the file name, the variable name, the exact change I wanted.
I had been making wishes the way people make wishes in fairy tales: loosely, assuming good faith interpretation, trusting the Genie to care about my wellbeing. But the Genie does not care. The Genie executes.
After the incident, I started cataloguing my bad wishes. Every ambiguous prompt I had written was a monkey's paw — a wish that gets granted in a way that technically satisfies the words but misses the point entirely.
"Refactor this function" — what does refactor mean? Rename variables? Extract helpers? Flatten the nesting? Change the signature? The model picks one, or picks all of them, or asks clarifying questions that eat tokens. I never specified.
"Make the output cleaner" — cleaner how? Shorter? Better formatted? Fewer warnings? The model guesses. Sometimes it guesses right. Often it does not, and then I spend more tokens correcting the correction.
"Write a blog post about X" — what length? What audience? What angle? What tone? What I am not allowed to say? What must be included? The model produces something generic because generic is the safest interpretation of an underspecified wish.
Every one of these is a monkey's paw wish. I wrote them and then felt frustration when the output missed the mark — but the frustration was misplaced. The Genie executed perfectly. I just made a terrible wish.
Precise wishes look different. They include everything the Genie needs to execute your intent, not just your surface request. They specify the context, the constraints, the format, the persona of the responder, the task in clear operational terms. They leave nothing for the Genie to infer.
I started thinking about what a precise wish actually requires. I kept arriving at the same list of missing ingredients every time I diagnosed a bad prompt. Something about who was speaking. Something about the surrounding context. Something about what was off limits. Something about what the output should look like.
I formalized this as the sinc format — six frequency bands that together fully specify a prompt the way the Nyquist theorem says you need at least 2x the signal frequency to reconstruct it without aliasing. Underspecify a prompt and the model hallucinates the missing parts. That hallucination is specification aliasing.
The six bands are PERSONA, CONTEXT, DATA, CONSTRAINTS, FORMAT, and TASK. Each one carries a measured portion of output quality. CONSTRAINTS carries 42.7% — by far the most. This is not an accident. CONSTRAINTS is where you stop hoping the model will infer your limits and start writing them down. It is the band where you specify what must not happen, what length is required, what assumptions are wrong, what the model will be tempted to do that you do not want.
A Genie with no constraints is terrifying. A Genie with explicit, exhaustive constraints becomes a precision instrument.
The formula above is not decoration. It describes why a missing band cannot be compensated by a stronger band elsewhere. The bands multiply. If CONSTRAINTS is empty, the product is zero no matter how good the TASK description is. The Genie hears a wish with a hole in it and fills the hole however it wants.
I stopped treating prompts as conversation starters. I started treating them as formal specifications — the same way I would write a function signature or a contract. The prompt is the entire interface. There is no room for implied context.
I built the sinc format into sincllm.com as AI Transform — a feature that takes a raw prompt and decomposes it into the six bands automatically, using a locally fine-tuned Qwen2.5-7B model running at 290 tokens per second on my RTX 5090. Zero API cost. The model makes the same decomposition I would make manually, in under a second, every time.
What I found across 275 observations: structured prompts reduce token cost by 97% and improve output quality to the point where the first response is almost always the final response. The Genie stops exploring because the wish is precise. The Genie executes and stops.
I still think about that 80,000-token incident. Not with frustration anymore — with gratitude. It was the most expensive lesson I have paid for in this space, and it was worth every token. It forced me to stop projecting human understanding onto a system that has no use for it, and start writing wishes the way a Genie requires them: complete, explicit, and exact.
The Genie is not going to read between your lines. Write the lines.
AI Transform decomposes any prompt into 6 sinc bands in under a second. Zero API cost. Built on a locally fine-tuned Qwen2.5-7B model.
Try AI Transform