Mario Alexandre · March 26, 2026 · sinc-llm structured-prompts prompt-engineering

The 6 Bands Every Prompt Needs (And What Happens Without Them)

I measured 275 prompt-response pairs. I scored which parts of a prompt drive quality. The results surprised me. The part most people focus on, the TASK description, matters least. The part almost nobody writes, CONSTRAINTS, matters most.

Here is the full picture. I will show what each band does and what happens when it is missing.

sinc-LLM — prompt as a 6-frequency-band signal

x(t) = Σ x(nT) · sinc((t - nT) / T)

n=0

PERSONA

Who the model should act like. It sets the expertise level, the voice, and what the model assumes you already know. Without it, the model picks a default that may be wrong for you. It might be too simple, too formal, or too careful.

7.0%

n=1

CONTEXT

What situation you are in. Your tech stack, your codebase, your business problem. Without it, the model answers in the abstract. It guesses about your setup and those guesses are often wrong.

6.3%

n=2

DATA

The facts, numbers, and examples that matter. Without them, the model has to guess the specifics. The chance of hallucination goes up sharply when DATA is empty.

3.8%

n=3

CONSTRAINTS

What the model must not do. Your rules, your limits, your hard requirements. Without them, the model can make any tradeoff it wants. It will often pick the wrong one. You will then spend several messages correcting it. Most back-and-forth comes from missing CONSTRAINTS.

42.7%

n=4

FORMAT

How the output should look. A code diff, a bullet list, a prose explanation, JSON, a step-by-step plan. Without it, the model picks a format on its own. It is often wrong. You end up with a 500-word essay when you wanted a diff.

26.3%

n=5

TASK

The actual ask. What you want done. Most people write only this part. Without the other 5 bands, TASK alone does not give the model enough to work with. The model can go wrong in a very convincing way.

2.8%

The Counterintuitive Finding

TASK carries 2.8% of the quality weight. CONSTRAINTS carries 42.7%. If you had to pick one to write perfectly, write the constraints section.

This makes sense when you think about it. The model is good at figuring out what you want. That is what it was trained to do. What it cannot do well is read your hidden rules. "Don't change the database schema" is not implied by any task description. "Keep the response under 200 tokens" is not something a model can guess. "Preserve the existing test coverage" is assumed by some models and ignored by others.

Every constraint you leave out is a possible wrong turn. Wrong turns create clarification questions, corrections, and do-overs. Those are your wasted tokens.

What Happens When Bands Are Missing

My measurement tracks the link between missing bands and exchange rate, meaning back-and-forth per prompt. The data is clear:

Missing CONSTRAINTS: the exchange rate goes up by about 1.8 extra responses per prompt. That is the single biggest driver.
Missing FORMAT: +1.1 extra responses per prompt.
Missing CONTEXT: +0.6 extra responses per prompt.
Missing PERSONA: +0.3 extra responses per prompt.
Missing DATA or TASK: almost no impact. The model infers or asks quickly.

That adds up fast. A prompt missing CONSTRAINTS, FORMAT, and CONTEXT burns about 3.5 extra round trips on average. At scale, 21,194 prompts per week, that is $1,588.56 in a week.

The auto-scatter hook fills all 6 bands for every prompt. You do nothing extra. The hook costs $0.002 per call. It saves $0.08 per call in avoided exchanges. That is the 38x ROI.

// Production AI Engineering

Build AI systems that hold up in production.

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 →