Skip to content
sinc-LLMBook a 30-minute call

Engineering reference

AI Transform: Model-Based Prompt Decomposition

A walkthrough of template and model-based prompt transformation, with before-and-after examples and the sinc prompt bands.

Methods, definitions, and examples for the topic discussed here. Check applicability against the task and operating conditions.

March 25, 2025 · 6 min read · ai-transform prompt-engineering sinc-llm local-llm
x(t) = Σ x(nT) · sinc((t − nT) / T)
Nyquist-Shannon sampling applied to prompt engineering. AI Transform samples your raw intent at 6 frequency bands, reconstructing it as a complete specification.

Two buttons, two approaches

sincllm.com now has two Transform buttons. "Transform" runs the original client-side template engine. It is fast, offline-capable, and has zero latency. "AI Transform" sends your raw prompt to a fine-tuned local model. That model reads your actual intent and produces sinc JSON tailored to what you are doing.

They produce the same output format. The difference is in how that output is made, and how good it is. When I first saw the AI output next to the template output side by side, the gap was obvious right away.

Transform (template)

Runs in the browser. Parses keywords and fills generic band templates. It is fast, but it only matches patterns. The output describes the format, not your intent.

AI Transform (model)

Calls a fine-tuned Qwen2.5-7B model running locally. It reads your full prompt and generates task-specific content in each band. It takes 1-2 seconds with zero marginal cost.

What templates can't do

A template engine matches the input to a pattern and slots tokens into set positions. For example, "fix the bug" might match a "debugging" template. That template fills PERSONA with "developer", TASK with "fix the identified bug", and CONSTRAINTS with a generic list of debugging constraints.

That is useful as a starting point. But it cannot do three things that matter:

It can't read specificity. "Fix the auth bug before the demo" and "Fix the rendering bug on mobile Safari" both match "debugging". But the CONTEXT, CONSTRAINTS, and TASK bands for these two prompts are completely different. A template produces the same output for both.

It can't infer domain. "Optimize the query" could be SQL, MongoDB, GraphQL, or a search engine query. The model looks at surrounding text and word choice to figure out which domain fits. Then it fills the PERSONA and CONTEXT bands correctly.

Review the proposed CONSTRAINTS field for task-specific rules and exclusions rather than accepting boilerplate.

Example 1: "fix the login bug"

Template output (CONSTRAINTS band):

"x": "Fix the issue correctly. Test thoroughly. Follow best practices.
Do not break existing functionality. Document changes."

AI Transform output (CONSTRAINTS band):

"x": "The login bug must be diagnosed before any fix is written —
read the error trace, identify the exact failure point (auth token
validation, session expiry, cookie handling, or redirect logic).
Fix only the broken component, not the login flow as a whole.
Existing authenticated sessions must not be invalidated by the fix.
The fix must be testable with a single test user without affecting
production sessions. No config changes that require a server restart
unless unavoidable. Regression test: login, logout, and re-login
must all succeed after the fix is applied."

The displayed structured output includes session-preservation, production-safety, and test requirements. Review whether those requirements match the task before using the prompt.

Example 2: "write a blog post about my new feature"

Template output (PERSONA band):

"x": "Content writer with expertise in the relevant domain."

AI Transform output (PERSONA band):

"x": "Technical product writer who understands both the engineering
implementation and the end-user value proposition. Writes in a
direct, specific voice — no fluff, no hedging. Uses exact metrics
and real examples. Treats readers as intelligent adults who want
information, not marketing copy."

The template produces a role. The model produces a writing philosophy and a voice specification that actually guides the output.

The six sinc bands

Band Function
PERSONA (n=0) Who is executing this. Role, expertise level, voice, approach.
CONTEXT (n=1) Situational framing. What is happening, why, and what system is involved.
DATA (n=2) Inputs, facts, references. What the executor is working with.
CONSTRAINTS (n=3) Rules, limits, invariants. What cannot be violated. Always longest.
FORMAT (n=4) Output shape. Length, structure, style, delivery format.
TASK (n=5) The atomic action. One sentence. What exactly is to be done.

Why CONSTRAINTS is always longest

Review the generated constraints against the task. Field length or a token allocation does not establish compliance in the output.

Think of it through Nyquist-Shannon: CONSTRAINTS is the sampling rate. If your sampling rate is too low, you get aliasing. The output folds into itself and produces artifacts (hallucinations, off-scope responses, wrong format). High-frequency CONSTRAINTS text prevents aliasing.

For a short input such as “hi,” inspect whether the generated fields are relevant to the greeting and appropriately concise. Check the CONSTRAINTS field for task-specific rules rather than assuming its length establishes quality.

Use the standard Transform when:

  • Your task requires a latency bound that the model-based transform has not been measured to meet.
  • You are experimenting with the format and do not need high-quality bands
  • The AI Transform endpoint is offline (local machine down)

Use AI Transform when:

  • You are generating a prompt that will run against a capable model
  • The task is domain-specific and template bands would be generic
  • You want to use the sinc JSON directly as a system prompt or agent instruction
  • You care about the CONSTRAINTS band being operationally correct

The two buttons coexist permanently. The template engine is the offline fallback. The model is the primary path. Both output the same JSON structure. The downstream consumer does not need to know which one ran. I designed it this way so there is never a dead end for the user.

// 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 →
Scope and context

sinc-LLM About Blog Services