Preprocessing vs Prompting

— The essence is preprocessing, not prompt design

Historically, prompt writing has been emphasized. AIKernel's position is clear:

Prompts are the final formatting step; inference quality is determined by preprocessing.

1. Why prompt design has been misunderstood

Many practitioners assume:

• Prompt wording is the key → ❌
• Few-shot examples improve accuracy → ❌
• Passing RAG as-is is acceptable → ❌

Attention Pollution Theory refutes these assumptions.

The fallacy of Few-shot

Few-shot is not a magic accuracy booster; it pulls attention to surface structures and poisons inference.

AIKernel treats Few-shot as an inference-breaking factor.

Passing RAG raw breaks structure

Raw data contains noise and pushes abstract structure out of attention.

2. The real question: what to put into attention

AIKernel focuses on controlling attention, not prompt style.

Put into attention:

• purpose
• constraints
• abstract structure
• reasoning patterns

Isolate from attention:

• examples
• stylistic instructions
• RAG fragments
• history
• noise

AIKernel deals with information structure, not just text.

3. Role of Preprocessing

— Deterministic control for OS-level stability

Preprocessing guarantees three OS-level properties:

3.1 Category separation

(See: Principles of Information Category Separation)

3.2 Context isolation

(See: Context Isolation Specification)

3.3 RAG structuring (Quarantine)

Treat raw RAG as potentially toxic material:

1. Retrieve (RAG / Provider)
2. Normalize (remove irrelevant data)
3. Structure (decompose, abstract, unitize)
4. Transfer only necessary parts to OrchestrationContext

This differs fundamentally from "throw RAG at the LLM."

4. Limits of Prompting

— Do not rely on nondeterministic processing

Prompts are nondeterministic LLM inputs:

• style
• examples
• explanations
• metaphors

These belong to ExpressionContext and must not affect inference.

You cannot reliably improve inference by prompt tweaks

Inference quality is set by preprocessing. Prompts only shape appearance.

5. Preprocessing is design, not craft

— Required to scale AI development

Prompt engineering often relies on intuition. Preprocessing is data-structure design and supports reproducibility, extensibility, and automation.

AIKernel emphasizes:

• structure over intuition
• design over craft
• categories over prose
• determinism over feel

This moves AI development from artisanal to engineering.

6. Conclusion

— Prompts are final formatting; inference quality is set by preprocessing

AIKernel's stance:

• Prompts belong to ExpressionContext
• Inference quality is determined by Preprocessing
• Preprocessing = category separation, structuring, isolation, quarantine
• AIKernel's essence is Preprocessing over Prompting

7. Positioning in AIKernel

This principle underlies:

• category separation
• context isolation
• attention pollution theory
• surface-mode prevention
• preprocessing-first design
• deterministic replay

AIKernel maximizes inference purity through information-structure design.

8. Related Documents

• Principles of Information Category Separation
• Context Isolation Specification
• Attention Pollution Theory
• Risks of Surface-Mode Failure
• AIKernel vs LangChain — Structural Comparison

Changelog

• v0.0.0 / v0.0.0.0: Initial draft
• v0.0.1 (2026-05-06): Version upgrade aligned with documentation guidelines