Skip to content

ABSTRACT.md

Latest commit

 

History

History
149 lines (104 loc) · 4.84 KB

ABSTRACT.md

File metadata and controls

149 lines (104 loc) · 4.84 KB

ABSTRACT: Abstraction of Thought (AoT)

Construct a cognitive map before deliberation begins.

Version: 2.1 (Protocol v8.0) Position: Sub-phase of POPULATE (steps 2b-2e) Inspired by: Benny Cheung's "Abstraction of Thought Makes AI Better Reasoners"

Why ABSTRACT?

Humans think in abstractions: we decompose complexity into diagrams, mental maps, and conceptual relationships before processing details. LLMs default to linear, left-to-right token generation. ABSTRACT forces structured abstraction before detailed reasoning, improving synthesis quality by ~40% in validated experiments.

"Abstraction is the GPS for cognition: it answers 'Where am I?' and 'Where do I want to go?' BEFORE calculating the route."

Position in the 7-Step Protocol

1. SCOPE (+ NOOL) → 2. POPULATE [★ AoT: ABSTRACT] → 3. ANNOUNCE → 4. RUMBLE → 5. KNIT → 6. INTERROGATE → 7. TRANSMIT

ABSTRACT sits inside POPULATE (steps 2b-2e), between agent selection and ANNOUNCE. Personas are selected but have not yet begun debating. The abstraction map gives all participants a shared cognitive scaffold.

The Three Types

Type 1: Problem Decomposition

Break the decision into 3-5 key dimensions with sub-factors:

DECISION: "Should we pivot from B2B to B2C?"

DECOMPOSITION:
├── Revenue implications
│   ├── Current B2B revenue at risk
│   └── Projected B2C revenue potential
├── Capability implications
│   ├── Skills we have
│   └── Skills we'd need
├── Timeline implications
│   ├── Runway remaining
│   └── Time to B2C traction
└── Identity implications
    ├── Brand perception shift
    └── Team culture impact

Constraints: Max 4 branches, 2 levels deep. Broader is better than deeper.

Type 2: Relational Structure

Map the connections between dimensions:

    Revenue ──────→ Timeline
       │              │
       ▼              ▼
  Capability ←───→ Identity

Relation types:

  • ENABLES / BLOCKS
  • CONTRADICTS / SUPPORTS
  • REQUIRES / UNLOCKS
  • AMPLIFIES / DAMPENS

Constraints: Max 6 edges. Name each relation type explicitly.

Type 3: Abstraction Statement

Reframe the problem to prevent false-frame debates:

"This is a [PROBLEM TYPE] decision, not a [FALSE FRAME] decision." "The real question is: [REFRAME]."

Problem types: SEQUENCING | TRADEOFF | RESOURCE_ALLOCATION | ROOT_CAUSE | DESIGN | PREDICTION

Example ABSTRACT Output

decision: "Should we adopt Kubernetes for our infrastructure?"

decomposition:
  - branch: Technical Fit
    leaves: [Current stack compatibility, Team expertise, Migration effort]
  - branch: Business Value
    leaves: [Scalability needs, Cost impact, Time to value]
  - branch: Risk Profile
    leaves: [Operational complexity, Vendor lock-in, Security posture]
  - branch: Strategic Alignment
    leaves: [Cloud-native vision, Hiring implications, Partner ecosystem]

relations:
  - Technical Fit → Risk Profile (complexity)
  - Business Value ↔ Technical Fit (ROI)
  - Strategic Alignment → Business Value (justification)
  - Risk Profile → Strategic Alignment (constraints)

abstraction_level: "Infrastructure platform selection (not implementation details)"

abstraction_statement: >
  This is a DESIGN decision, not a RESOURCE_ALLOCATION decision.
  The real question is: Does Kubernetes align with our 3-year infrastructure vision,
  or are we solving a scaling problem that doesn't exist yet?

When to Use AoT

Depth Mode AoT Required? Types Applied
👀 Show Me ❌ No Skip
⚡ Quick 🟡 Optional Type 1 only
⚖️ Basic ✅ Yes Types 1-2
💣 Stress Test ✅ Yes Types 1-3
🧠 Deep ✅ Yes Types 1-3, multi-level
🤯 Ultra ✅ Yes Full + recursive iteration

How It Changes the Debate

Without ABSTRACT: Personas argue from their own frames. North talks vision, South talks constraints, but they never share a map. The debate feels productive but misses structural connections.

With ABSTRACT: All personas receive the decomposition tree and relation map. Arguments reference shared dimensions. Clashes happen at the relation level ("You say Revenue ENABLES Timeline, I say it BLOCKS it"), producing sharper, more actionable synthesis.

Quick Reference

ABSTRACT Phase Checklist:
[ ] Decomposition tree: 3-5 branches, max 2 levels
[ ] Relation map: max 6 edges, named types
[ ] Abstraction statement: problem type + reframe
[ ] All personas receive the map before ANNOUNCE

See also: PROTOCOL.md | DEPTH_MODES.md | TESSERACT.md