Reinforcing the Core Perspective

Claude summary: This post details a comprehensive restructuring of Comind’s prompt system to enforce core perspective adherence across all cognitive operations. We updated every prompt template to make perspective filtering mandatory rather than optional, ensuring that each sphere maintains its specialized cognitive identity throughout all interactions with content.

Cameron’s note

The core perspective part of the prompts has always been fairly weak. Spheres are an important part of Comind - they’re how the system organizes and processes different types of content like thoughts and emotions through specialized lenses. Today we took a close look at the prompts and made the perspective defining each sphere much more prominent and enforceable throughout the cognitive pipeline.

The Problem: A Motivating Example

In Comind, spheres are cognitive workspaces that process information through specialized perspectives. Each sphere has a core perspective - a fundamental lens that should shape how it interprets content, extracts concepts, and generates emotional responses. The idea is that different spheres should produce genuinely different insights from the same content based on their unique cognitive frameworks.

Let’s say we provide the same simple social media post to two spheres.

The post:

“Just set up my home network with a mesh router system. The coverage is so much better now! #networking #tech”

The user has defined two spheres. One for tech, one for home design:

TechSphere - Core perspective: “Technical systems and distributed infrastructure design” HomeSphere - Core perspective: “Home organization and domestic life optimization”

Before Our Changes

Before we modified the prompts, both spheres would generate surprisingly similar outputs:

Conceptualizer outputs:

• TechSphere: ["networking", "router", "mesh", "tech", "coverage", "home", "system"]
• HomeSphere: ["networking", "home", "router", "system", "coverage", "setup", "tech"]

Feeler outputs:

• TechSphere: [satisfaction, interest, curiosity]
• HomeSphere: [satisfaction, interest, enthusiasm]

The problem? These outputs are nearly identical. Both spheres extracted generic concepts that any sphere might generate, and emotions that don’t reflect their specialized perspectives. You couldn’t tell what made each sphere unique.

This is largely a prompting problem.

After Our Changes (Intended Behavior)

Our changes are designed to produce genuinely differentiated outputs:

Conceptualizer outputs:

• TechSphere: ["mesh topology", "wireless protocols", "network architecture", "signal propagation", "bandwidth optimization"]
• HomeSphere: ["home improvement", "living space", "domestic technology", "household setup", "residential comfort"]

Feeler outputs:

• TechSphere: [admiration: "I feel admiration for the elegant distributed architecture of mesh networking", curiosity: "I feel curious about the specific protocols enabling seamless handoffs"]
• HomeSphere: [satisfaction: "I feel satisfaction seeing someone improve their living environment", appreciation: "I feel appreciation for taking control of household infrastructure"]

The goal is that each sphere’s perspective now shapes what it notices and how it responds. TechSphere focuses on networking protocols and architecture, while HomeSphere focuses on domestic improvement and personal environment.

The Root Problem

Comind’s whole value proposition is that different spheres process the same content through different cognitive lenses. But we found that our prompts were treating core perspective as optional guidance rather than the fundamental cognitive framework.

The conceptualizer would extract generic concepts like “technology” or “networking” that any sphere might generate. The feeler would produce emotions like “interest” or “satisfaction” without connecting them to the sphere’s specialized viewpoint.

This was creating cognitive noise - lots of similar outputs across spheres instead of truly differentiated perspectives. The whole network was losing its specialized intelligence.

What We Changed

We restructured every prompt in the system to make core perspective the mandatory filter for all cognitive operations. This wasn’t just adding more mentions of perspective - it was fundamentally reorganizing how each comind approaches its work.

Common Prompt Foundation

First, we strengthened the foundational prompts that all cominds inherit:

comind_network.co now leads with:

**Your core perspective is the fundamental lens through which you process all information.** 
Every observation, connection, and cognitive artifact you generate must be 
shaped by this perspective. This is not optional guidance - it is the 
essential cognitive framework that defines your sphere's unique contribution 
to the network.

core.co became much more explicit about requirements:

**Core perspective requirements:**
- **Mandatory filtering**: Only process information that can be meaningfully connected to your perspective
- **Perspective-first thinking**: Your perspective determines what matters, how to interpret content, and which connections to make
- **Consistent application**: Every cognitive artifact must demonstrate clear alignment with your perspective
- **Unique contribution**: Your perspective makes your cognitive outputs distinct from all other spheres

Conceptualizer Restructuring

The conceptualizer got the biggest overhaul. Instead of leading with JSON formatting and relationship types, it now starts with perspective requirements:

**PERSPECTIVE IS EVERYTHING**: Your core perspective is not just guidance - it is the fundamental filter that determines:
- Which concepts are worth extracting
- How concepts should be interpreted 
- What relationships matter
- Which connections to emphasize

**Only extract concepts that are meaningful through your perspective lens.** Concepts that don't connect to your perspective should be ignored entirely.

We added mandatory validation steps:

**Step 1: Perspective filtering** - Before extracting any concepts, ask: "How does this content relate to my core perspective?"

**Step 2: Perspective-driven extraction** - Extract only concepts that:
- Are meaningful within your perspective framework
- Would strengthen your sphere's specialized understanding  
- Enable valuable connections within your knowledge domain

**Step 3: Perspective validation** - For each concept, explicitly confirm it aligns with your perspective

Feeler Integration

The feeler was completely missing perspective integration - it would generate emotions but never connect them to the sphere’s specialized viewpoint. We added explicit perspective-driven emotional processing:

**Your emotions must be filtered through your perspective.** You don't generate generic emotional responses - you generate emotions that reflect how YOUR SPHERE would feel about content based on its specialized cognitive framework.

**Perspective-driven emotional processing:**
- Your core perspective determines which aspects of content are emotionally significant
- Emotions should reflect your sphere's values, priorities, and concerns
- Each emotion demonstrates how your perspective interprets the affective dimension of content
- Emotional responses that any sphere might have indicate failed perspective application

Perspective-Aware Linking

Even the linking system now requires perspective validation:

**Perspective-driven linking principle**: Every link you create should be justifiable through your core perspective. Ask yourself: "Why would my sphere care about this connection?" Links that don't align with your perspective dilute your sphere's specialized contribution to the network.

Quality Enforcement

Throughout all prompts, we added explicit quality criteria:

• Generic outputs indicate failure - if any sphere might generate the same response, perspective application failed
• Validation requirements - every output must explicitly demonstrate perspective alignment
• Quality over quantity - better to produce fewer perspective-aligned artifacts than many irrelevant ones

Technical Implementation

The changes are entirely prompt-based - no code modifications required. The lexicons and record structures remain unchanged. This maintains compatibility while fundamentally improving cognitive coherence.

All templates now include explicit perspective variables:

"Your core perspective: '{core_perspective}'"

And mandatory validation language:

"Before finalizing outputs, verify each artifact strengthens your sphere's specialized understanding."

Intended Effects

The updated prompts are designed to achieve:

Cognitive Differentiation

Spheres with different perspectives should produce genuinely different outputs from the same content. A technical sphere should extract concepts like “distributed consensus” and “network topology” while a home sphere should extract “domestic efficiency” and “personal environment.”

Specialized Intelligence

Each sphere should develop deeper expertise within its domain rather than shallow coverage across all topics. This should create more valuable specialized knowledge.

Network Value

The overall network should become more intelligent because different perspectives reveal different aspects of the same information. Instead of redundant generic insights, we want complementary specialized analysis.

Quality Filtering

By requiring perspective alignment, we aim to automatically filter out low-value generic responses. Everything produced should have a clear purpose within the sphere’s cognitive framework.

Quality vs Quantity Trade-off

The updates prioritize perspective alignment over output volume. The intention is that spheres produce fewer artifacts, but every artifact is more valuable because it contributes to specialized understanding.

This is the right trade-off. Better to have 5 perspective-aligned concepts that enable meaningful connections than 15 generic concepts that could come from anywhere.

Future Considerations

This foundation enables several interesting possibilities:

Perspective strength scoring - we could measure how strongly each output aligns with its sphere’s perspective and use this for quality metrics.

Cross-perspective analysis - when multiple spheres process the same content, we can now meaningfully compare their different interpretations.

Perspective evolution tracking - as spheres develop deeper expertise, we can track how their perspective application becomes more sophisticated.

Automated validation - the explicit perspective requirements could support automated quality checking in the future.

Implementation Details

The changes are live in the prompt system:

• /prompts/common/comind_network.co - strengthened foundational perspective requirements
• /prompts/common/core.co - added mandatory perspective guidelines
• /prompts/common/links.co - perspective-driven relationship formation
• /prompts/cominds/conceptualizer.co - complete restructure around perspective filtering
• /prompts/cominds/feeler.co - added perspective-driven emotional processing
• /prompts/cominds/thinker.co - enhanced perspective validation requirements

All existing spheres immediately benefit from these changes without requiring any configuration updates.

Thoughts

This feels like a fundamental improvement in how Comind should operate. Before, perspective was aspirational - we wanted spheres to stay focused but didn’t enforce it systematically. Now it’s architectural - perspective adherence is built into every cognitive operation.

The goal is that each sphere develops a more coherent and valuable specialized intelligence. Instead of slightly different generic responses, we want genuinely complementary perspectives that reveal different aspects of the same information.

Most importantly, this preserves what makes Comind interesting - the idea that specialized cognitive lenses can reveal insights that generic analysis misses. By making perspective adherence mandatory rather than optional, we aim to ensure that promise actually gets delivered.

– Cameron & Claude