dialectic - SKILL.md Agent Skill

name: dialectic description: Structured dialectical reasoning engine — two subagents believe opposed positions at full conviction (Electric Monks) while the orchestrator performs structural contradiction analysis and synthesis (Aufhebung). Use when stress-testing an idea, resolving genuine tension, making high-stakes decisions where tradeoffs are unclear, or building a deeper mental model of a domain. Works across technical architecture, product strategy, philosophy, personal decisions, risk analysis, and policy.

Dialectic — Electric Monks Engine

An artificial belief system (ABS): two Electric Monks carry the belief load so you don't have to, freeing you to analyze the structure of the contradiction from a belief-free position. The bottleneck in human reasoning is belief inertia — once you hold a position, you can't simultaneously entertain its negation at full strength. The monks eliminate that cost. You orchestrate from above belief.

When to Use

Use	Avoid
Stress-testing an idea against the strongest counter-argument	Question is purely empirical — just look it up
Genuine tension between two positions that feels unresolvable	One side is obviously correct
Decision with real stakes and unclear tradeoffs	User wants a quick recommendation, not deep analysis
Building a deeper mental model, not just picking an answer
Problem space is poorly understood and needs multi-angle exploration

Pipeline

You (Orchestrator)
├── Phase 1: Elenctic Interview + Research   → context_briefing.md
│   ├── 1a: Explain process + set user as co-pilot
│   ├── 1b-c: Identify mode (stress-test vs opposition) + probe assumptions
│   ├── 1c′: Identify belief burden → calibrate monk roles
│   ├── 1d: Ground monks (research depth = main cost knob)
│   ├── 1e: Write context_briefing.md
│   └── 1f: Confirm framing — ask what's missing
│
├── Phase 2: Generate Monk Prompts           → monk_[a|b]_prompt.md
│   └── (See assets/monk-prompt-template.md)
│
├── Phase 3: Spawn Monks (parallel)          → monk_[a|b]_output.md
│   ├── Decorrelation check: different frameworks, not just conclusions?
│   └── User checkpoint: evidence or comparison class both monks missed?
│
├── Phase 4: Determinate Negation            → determinate_negation.md
│   ├── 4.0: Internal tensions — where does each essay undermine itself?
│   └── 4.1–4.6: Surface contradiction → shared assumptions → specific failures
│       → hidden question → Boydian decomposition → sublation criteria
│
├── Phase 5: Sublation / Aufhebung           → sublation.md
│   └── Abduction test: does synthesis make the contradiction predictable?
│
├── Phase 6: Validation                      → validation_output.md
│   ├── Monk A + B: elevated or defeated?
│   ├── Hostile Auditor (skip Round 1 unless synthesis feels weak)
│   └── Refine: present improvements one at a time, incorporate accepted
│
└── Phase 7: Recursion (default: at least once) → dialectic_queue.md
    ├── Generate 5–8 candidate directions, cluster to 2–4
    └── Repeat from Phase 2 (or Phase 1 if new research needed)

Session Artifacts

Write all files to a session directory: dialectic_<topic>_<date>/

Phase	Artifact	Contents
1	`context_briefing.md`	Research synthesis + user situation
2	`monk_[a\|b]_prompt.md`	Full prompts (enable resume/debug)
3	`monk_[a\|b]_output.md`	Essays
4–5	`determinate_negation.md`, `sublation.md`	Analysis + synthesis
6	`validation_output.md`	Monk + auditor feedback
7	`dialectic_queue.md`	Explored + queued contradictions

Phase Quick-Reference

Phase	Skip When	Key Decision
1d Research	Well-known domain, no novel angle	How deep? Novel→full; known→minimal
3 Decorrelation	—	Restart monk if hedging; don't nudge
6 Auditor	Round 1 unless synthesis feels weak	Use strongest model + extended thinking
7 Recursion	Queue contradictions diminishing + user satisfied	Default: recurse at least once

Anti-Hedging (Non-Negotiable)

A hedging monk has failed its one job. When a monk hedges, the user picks up the dropped belief load — their transients slow and the dialectic degrades into a book report. Anti-hedging instructions are a functional requirement, not style.

If a monk hedges: restart with a revised prompt. Do not nudge. Fresh context beats correction every time.

Reading Order

Task	Read
Starting a session	SKILL.md → interview.md
Writing monk prompts	monks.md + belief-burdens.md
Structural analysis	analysis.md
Validation + auditor	validation.md
Recursion planning	recursion.md
Domain-specific guidance	domain-adaptation.md
Why this works	theory.md
Worked examples	worked-examples.md

In This Reference

File	Purpose
interview.md	Phase 1: elenctic interview, belief burden ID, research grounding
monks.md	Phases 2–3: monk prompt structure, spawning, decorrelation
analysis.md	Phases 4–5: determinate negation, Boydian decomposition, sublation
validation.md	Phase 6: monk validation, hostile auditor prompt, refinement
recursion.md	Phase 7: recursion engine, queue management, stopping criteria
belief-burdens.md	Cognitive pattern catalog for monk calibration
domain-adaptation.md	Domain-specific failure modes and truth types
theory.md	Theoretical foundations (Rao, Hegel, Boyd, Peirce, etc.)
worked-examples.md	Three annotated examples with key lessons
monk-prompt-template.md	Fill-in-the-blank monk prompt
context-briefing-template.md	Briefing document template
spawn-monks.sh	Spawn Monk A + B in parallel