Cognitive Scaffolding

# Cognitive Scaffolding

Part of [Agent Skills™](https://github.com/itallstartedwithaidea/agent-skills) by [googleadsagent.ai™](https://googleadsagent.ai)

## Description

Cognitive Scaffolding structures an agent's context window using principles from cognitive science — primacy effects, recency bias, chunking, and attention allocation. Language models, like human working memory, are not uniform processors. Information placed at the beginning and end of the context receives disproportionate attention (primacy and recency effects), while content in the middle can be effectively invisible. Cognitive Scaffolding exploits these properties to ensure the most critical information receives maximum model attention.

This skill was developed through extensive experimentation on the Buddy™ agent at [googleadsagent.ai™](https://googleadsagent.ai), where analysis accuracy improved by measurable margins simply by restructuring how information was arranged in the context window. Campaign performance data placed at strategic positions within the prompt produced significantly better recommendations than the same data placed arbitrarily. The same principle applies to code context, documentation, and any other information an agent must reason over.

The cognitive scaffolding framework organizes context into four zones: the anchor zone (first 5% of context — highest attention, used for identity and immutable rules), the foreground zone (last 20% — high attention, used for the current task and recent context), the structured middle (60% — moderate attention, organized into clearly delimited chunks), and the background zone (15% — lowest attention, used for reference material and fallbacks). Each zone has specific content strategies that maximize the model's ability to utilize the information placed there.

## Use When

- Agent accuracy varies inconsistently despite using the same information
- Long context windows degrade performance compared to shorter interactions
- Critical instructions or constraints are occasionally ignored by the agent
- You need to present large amounts of reference data without overwhelming the agent
- Multi-document reasoning requires the agent to attend to specific sections
- You are optimizing agent behavior for specific model architectures

## How It Works

```mermaid
graph LR
    subgraph "Context Window"
        A[Anchor Zone<br/>5% - Highest Attention<br/>Identity, Rules] --> B[Structured Middle<br/>60% - Chunked Data<br/>Delimited Sections]
        B --> C[Background Zone<br/>15% - Reference<br/>Fallback Material]
        C --> D[Foreground Zone<br/>20% - Current Task<br/>Recent Messages]
    end
    
    E[Primacy Effect] -.-> A
    F[Chunking Theory] -.-> B
    G[Low Attention Region] -.-> C
    H[Recency Effect] -.-> D
```

The scaffolding exploits well-documented attention patterns in transformer architectures. The anchor zone leverages primacy — the model attends strongly to the earliest tokens, making this the ideal location for identity statements and non-negotiable rules. The foreground zone leverages recency — the most recent tokens receive high attention, making this the best place for the current task and recent conversation. The structured middle uses explicit delimiters and headers to create navigable chunks that compensate for the "lost in the middle" effect. The background zone stores material the model can reference but doesn't need to actively attend to.

## Implementation

**Cognitive Zone Builder:**

```typescript
interface CognitiveZone {
  name: string;
  position: "anchor" | "middle" | "background" | "foreground";
  budgetPercent: number;
  content: string;
  delimiter: string;
}

class CognitiveScaffold {
  private totalBudget: number;
  private zones: Map<string, Cogniti