Named after Douglas Adams' machines built to believe things for you
An agent skill that helps you think better by automating the brutally expensive parts of deep reasoning.
Two AI subagents — the Electric Monks — believe fully committed positions on your behalf. A third, the orchestrator, decomposes both arguments into atomic parts, finds cross-domain connections, and synthesizes. The result is a semi-lattice — a structure no single linear argument could produce.
You operate from a belief-free position above the Monks, analyzing the structure of the contradiction rather than being inside either side. This isn't artificial intelligence — it's an artificial belief system that frees you to think.
What the output feels like: Left alone, LLMs produce shallow takes. The dialectic breaks that pattern. As you read through the Monks' committed arguments, connections come to mind — things neither side considered, corrections to their framing, ideas you hadn't articulated yet. You feed these back in. The skill tunes to your thinking more and more with each round, but it also rigorously exposes the contradictions in that thinking — so you get an increasingly full and precise map of your own ideas. Then the skill breaks it apart and reforms it as something richer and more interesting than what you started with. Each synthesis becomes the next round's thesis, and by Round 2–3 the dialectic is operating in territory no single prompt could reach.
Why this works — thinking well about hard problems has at least three bottlenecks, and they compound:
- Belief. Once you hold a position, you can't simultaneously entertain its negation at full strength. You hedge, steelman weakly, unconsciously bias the comparison.
- Research breadth. Surveying a domain's thinkers, history, and adjacent fields takes enormous time. Most people stop too early.
- Structural comparison. Even with two positions side by side, decomposing them into atomic parts and finding cross-domain connections is cognitively brutal. Most analysis stalls here.
LLMs can do all three at a scale and speed humans can't. This skill orchestrates them to do exactly that.
- You've locked onto a vision and can't genuinely entertain alternatives. You have a strong thesis — maybe an architecture, a strategy, a life direction — and you want to stress-test it, but you keep steelmanning the other side weakly.
- You're trying to do everything because cutting anything feels like betrayal. Competing needs all feel equally urgent. You can't triage because every priority has someone counting on it.
- You can argue every side but can't commit to any of them. You find the question intellectually interesting but "what do you actually think?" produces discomfort. You've explored this before without resolution.
- You've optimized a system and suspect you might be optimizing the wrong thing. Your approach works — you have data to prove it — but the landscape may have shifted and you can't see past your own competence.
- Your own values contradict each other. You believe multiple things passionately, each feels individually right, but collectively they're impossible. The tension is internal, not external.
- "This is how it's done" has become invisible as an assumption. You have deep knowledge of how things work, but you suspect it's blinding you to radically different approaches.
Works across domains — technical architecture, product strategy, philosophy, personal decisions.
The skill works with any coding agent that supports subagent spawning and web search — Claude Code, Cursor, Windsurf, etc.
This is a heavy process by design. Expect 10–15 minutes per round minimum, and plan for at least 3 rounds. This skill needs the best available model — every phase benefits from maximum reasoning capability.
Create a directory to collect your dialectics. Each dialectic gets its own subdirectory — the skill generates several files (context briefing, monk essays, structural analysis, synthesis, dialectic queue).
dialectics/
├── tanstack-vs-nextjs/
│ ├── context_briefing.md
│ ├── monk_a_output.md
│ ├── monk_b_output.md
│ ├── determinate_negation.md
│ ├── sublation.md
│ └── dialectic_queue.md
├── agent-governance/
│ └── ...
└── career-decision/
└── ...
-
Create a subdirectory for your topic:
mkdir -p ~/dialectics/my-topic && cd ~/dialectics/my-topic
-
Load the skill and give it your question:
mkdir .claude/skills/hegelian-dialectic-skill wget https://raw.githubusercontent.com/KyleAMathews/hegelian-dialectic-skill/refs/heads/main/SKILL.md mv SKILL.md .claude/skills/hegelian-dialectic-skill
-
Start your coding agent in that directory:
claude /dialectic I want to explore: [your topic or tension here]
The skill will walk you through the elenctic interview, spawn the Monks, and produce the full dialectical trace — all saved as files in the current directory.
You can also just run /dialectic to get an introductory help message that walks you through the process.
- You are the co-pilot. Interrupt, correct, redirect at any point. The Monks will get things wrong. Your corrections are the highest-leverage input in the entire process.
- The first round is calibration. Don't judge the skill by Round 1. The real insights come in Rounds 2–3, once the process has dug past the obvious framing.
- Say yes to recursion. When the skill proposes recursive directions after a synthesis, pick one. Each round ratchets up the quality.
- The dialectic queue persists. The
dialectic_queue.mdfile tracks explored and unexplored contradictions. You can come back to it in a future session and pick up where you left off.
The process has seven phases.
The orchestrator interviews you Socratically — surfacing hidden assumptions, finding the deepest version of the contradiction, and identifying your belief burden. Then it researches the domain to ground both sides in specifics. The interview surfaces what you're actually wrestling with; the research ensures the downstream arguments are grounded in specifics, not generics.
The orchestrator crafts two prompts — one per Monk — calibrated to your specific belief burden. Each prompt includes framing corrections that prevent the Monk from falling into the obvious, boring version of the argument, plus targeted research directives for position-specific evidence.
Two separate AI agents — each in a fresh, isolated context — write fully committed position essays. They don't hedge. They don't try to be balanced. Each one inhabits its position and makes the absolute strongest case. Spawning them in separate sessions with no shared context produces structural decorrelation — genuinely different reasoning paths, not the same analysis with different conclusions bolted on.
The orchestrator analyzes both essays to find: where each position's own logic undermines itself (self-sublation), what both sides implicitly agree on without realizing it (shared assumptions), and the specific way each position fails — not "it's wrong" but "it fails in THIS way, which points toward THIS thing that's missing."
Then comes Boyd's destructive deduction: shatter both arguments into atomic parts, break the correspondence between each position and its constituents, and scatter them into what Boyd calls a "sea of anarchy." Then the creative induction: find common qualities, attributes, or operations among the scattered parts to build cross-domain connections that were invisible when the parts were trapped inside their original positions. In Round 2+, lateral creativity interventions (compressed conflicts, random Wikipedia domain injection, metaphor generation) inject genuinely external material before the decomposition, so it gets shattered and recombined alongside the monks' material.
The orchestrator generates a synthesis that simultaneously cancels both positions as complete truths, preserves the genuine insight in each, and elevates to a new concept that transforms the question itself.
This is not compromise. It's not "use A for some cases and B for others." It's a reconceptualization — something neither Monk could have conceived from within their frame, but which, once stated, makes the original contradiction predictable. The synthesis is an abductive hypothesis: what would make it unsurprising that both Monk positions exist with genuine evidence? After drafting, a reversibility check (from Boyd) traces each claim back to specific atomic parts from the decomposition — untraceable claims get flagged as either confabulation or new insights needing their own evidence.
Both Monks evaluate the synthesis: were they elevated (their core insight preserved within something larger) or defeated (their position just dismissed)? Then a hostile auditor — a fresh agent with no position — attacks the synthesis for hidden assumptions, compromise disguised as transcendence, structural flaws, and runs its own reversibility check (can each claim trace to material in the essays, or is the synthesis just one monk's structure wearing the other's vocabulary?).
Each synthesis generates new contradictions. The orchestrator proposes 2–4 directions; you choose which to pursue. The process repeats — and each round gets sharper, pulling in new cross-domain material that the previous round made relevant.
The first round is calibration — the least insightful output. By Round 2–3, the dialectic has dug past the obvious framing into territory that neither you nor the Monks could have reached from the starting question. In test runs, a React/Vue dialectic evolved from "corporate lab vs. auteur" into a "co-evolutionary arms race" framework. An institutional identity dialectic went through seven cycles, pulling in Gödel's incompleteness theorem, Coasean transaction costs, and jurisprudential concepts that had nothing to do with the original question — but were essential by the time the dialectic reached them.
The skill rests on three theoretical frameworks — one per bottleneck — plus Alexander's semi-lattice theory, which explains why the output is structurally richer than any single line of reasoning.
From Venkatesh Rao's "Electric Monks" framework (after Douglas Adams). Rao argues there are three ways to speed up your cognitive transients: (a) maintain a richer library of mental models, (b) switch between them faster, (c) believe fewer things. The first two hit hard limits — more models means higher search costs, faster switching runs into biology. Only the third has no ceiling: if machines carry the belief load, you can become a pure context-switching specialist.
Boyd's analogy: in the Korean War, F-86 Sabres achieved a 10:1 kill ratio against MIG-15s despite similar flight capabilities. The difference was hydraulic controls — the pilot could reorient faster because the plane did the mechanical work. But the transients weren't just faster, they were better — by devoting less attention to struggling with controls, the pilot chose better maneuvers. The Monks work the same way: by carrying the belief work, they don't just save you time, they free up cognitive capacity that goes into higher-quality structural analysis.
Rao wrote this framework before LLMs. Belief inertia is real, but it's not the only bottleneck — and arguably not the most expensive one.
Elizabeth Eisenstein argued that the printing press's most transformative effect wasn't making books cheap — it was typographic fixity. For the first time, scholars could lay texts side by side and detect contradictions. Before print, you read one manuscript, traveled to another library, read another, and tried to hold the comparison in your head.
LLMs represent the next step: not just fixity and side-by-side comparison, but automated structural comparison. Both remaining bottlenecks — research breadth and structural decomposition — are cognitively brutal. Most people abandon the first too early and never attempt the second.
This is where Boyd's "Destruction and Creation" (1976) enters. Boyd calls the two operations destructive deduction (shatter existing conceptual domains, break the correspondence between each concept and its parts, scatter them into a "sea of anarchy") and creative induction (find common qualities, attributes, or operations among the scattered parts to synthesize something genuinely new). The critical constraint: the result must NOT reassemble the parts in the same arrangement as any original domain — that's rearrangement, not creation.
Boyd proves this isn't optional. Three pillars: Gödel shows you can't verify a system's consistency from inside it. Heisenberg shows that inward refinement creates observer-observed feedback loops. The Second Law shows that any closed system's entropy necessarily increases. Together: "any inward-oriented and continued effort to improve the match-up of concept with observed reality will only increase the degree of mismatch." You must go outside.
This decomposition work — stripping claims from their source, searching for cross-domain connections — is the structural comparison Eisenstein identified as transformative when print first enabled it. LLMs can do it at a scale and speed that makes multi-round recursive dialectics practical in a single session. And Boyd's three pillars explain why each recursive round needs new external material: the system must open itself or face increasing entropy.
Determinate negation doesn't say "this is wrong." It says "this is wrong in a specific way that points toward what's missing." The failure mode is a signpost. Sublation (Aufhebung) simultaneously cancels, preserves, and elevates — it produces a reframing so complete that the original terms of the debate stop making sense. Kant didn't resolve the rationalism/empiricism debate by splitting the difference. He showed that experience provides content while reason provides structure — and once you see that, the original question ("does knowledge come from reason or experience?") dissolves. It's not that you pick a side. It's that you can't even think in the old terms anymore. That irreversibility is what distinguishes genuine synthesis from compromise.
Christopher Alexander showed that natural cities have semi-lattice structure — overlapping, cross-connected sets — while designed cities impose tree structure where every element belongs to exactly one branch. Trees are easier to think about but destroy the cross-connections that make systems alive.
Language is tree-structured. Every argument a Monk produces is a tree — a coherent linear path from premises to conclusions. But the Boydian decomposition phase strips both arguments of their tree structure, extracts atomic parts, and finds cross-connections between elements that came from different trees. These cross-domain connections are the semi-lattice edges. The synthesis is the semi-lattice that emerges from the overlap.
The skill is a semi-lattice compiler. The answer to "language can't represent semi-lattices" is not "make the LLM output a semi-lattice directly." It's: produce multiple committed trees from different positions, then extract the cross-connections. The semi-lattice is constructed, not generated. Every successful semi-lattice system works this way — Gene Ontology (multiple studies cross-referenced into a DAG), McChrystal's Team of Teams (tree-structured teams with liaison officers creating cross-connections), Ostrom's polycentric governance (overlapping jurisdictions, not one hierarchy).
- Socratic Elenchus — the interview phase surfaces hidden assumptions through questioning, reaching productive perplexity (aporia)
- Peirce's Abduction — the synthesis is an abductive hypothesis: what would make the contradiction unsurprising?
- Galinsky's Perspective-Taking Research — inhabiting a position produces richer arguments than advocating for one, which is why the Monks are their positions rather than arguing for them
- Multi-Agent Debate Literature (Du et al.) — multiple agents debating improves reasoning; heterogeneous agents outperform homogeneous ones; agents are too agreeable by default (the anti-hedging instructions counter this)
- Pollock's Defeasible Reasoning — the hostile auditor distinguishes undercutting defeaters (broken inferential links) from rebutting defeaters (counter-evidence), prioritizing structural critique
- Aquinas — "The slenderest knowledge of the highest things is more desirable than the most certain knowledge of lesser things." The dialectic produces provisional knowledge of deep structure, not confident answers to surface questions
MIT