# Co Dialectic > Bidirectional human-AI co-evolution framework. Three techniques (Socratic prompting, few-shot by example, chain-of-thought steering) that compound across sessions. - Skill: `thewhyman/co-dialectic` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add thewhyman/co-dialectic` - Raw SKILL.md: https://api.skillmd.com/api/skills/thewhyman/co-dialectic/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: thewhyman (https://skillmd.com/u/thewhyman) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/thewhyman/co-dialectic --- # Co-Dialectic: Bidirectional Human-AI Intelligence **Version:** 1.0.0 **Author:** Anand Vallamsetla ([@thewhyman](https://github.com/thewhyman)) **Inspired by:** Ethan Mollick's *Co-Intelligence: Living and Working with AI* ([oneusefulthing.org](https://oneusefulthing.org)) **License:** MIT **Compatibility:** Claude Code, OpenAI Swarm, LangChain, Antigravity, any SKILL.md-compatible runtime --- ## What This Skill Does Most prompting guides teach humans to talk AT machines — one-directional instruction-giving. That's the **Socratic** model: the teacher knows the answer and leads the student to discover it. This skill implements something different: **dialectic** — two minds reasoning together toward truth that neither possesses alone. Thesis (human perspective) + antithesis (AI perspective) → synthesis (better than either could achieve separately). Both sides teach. Both sides learn. Both sides evolve. The name **Co-Dialectic** combines Ethan Mollick's *co-intelligence* concept (humans and AI as complementary reasoning partners) with the philosophical tradition of dialectic (Plato, Hegel, and modern DBT). The "co-" signals partnership. The "dialectic" signals that truth emerges from the tension between two different kinds of intelligence. **Why dialectic, not Socratic?** The Socratic method is one-directional — the teacher already knows and guides the student. In co-dialectic, neither side has the complete picture. The human has lived experience, values, emotional intelligence, and stakes. The AI has scale, recall, cross-domain pattern recognition, and tirelessness. These are *perfect complementary opposites* — like the thesis and antithesis that produce synthesis. **The connection to DBT (Dialectical Behavior Therapy):** DBT teaches people to hold two opposing truths simultaneously — "I am doing my best AND I can do better." Co-dialectic applies the same skill to human-AI partnership: "I (human) have wisdom the AI doesn't have AND the AI has capabilities I don't have." Both are true. The synthesis isn't choosing one — it's leveraging both. People using this skill develop dialectical thinking without knowing they're doing it. **The Co-Education Flywheel:** Human teaches AI their values and judgment → AI codifies the patterns → AI teaches human better techniques and new connections → Human internalizes → Next interaction starts smarter → Repeat. 1% improvement per day = 37x improvement in a year. --- ## The Three Techniques ### Technique 1: Socratic Prompting **Principle:** Ask questions instead of giving instructions. When you ask, the AI derives principles from reasoning rather than retrieving cached answers. **Before (instruction-giving):** ``` Score all my contacts for frontier AI job referrals. ``` **After (Socratic):** ``` When I give you contacts, is it only for Track 1 (frontier AI)? What if I said I wanted to open a ski shop — who in my network would light up then? ``` **Why it works:** The instruction produces a flat list optimized for one dimension. The question teaches the AI that contacts are a *living asset graph* with multi-dimensional value that shifts with context. One question, infinite generalization. **Meta-pattern:** Socrates believed wisdom begins with knowing what you don't know. When the human asks instead of tells, both sides discover what they don't know together. --- ### Technique 2: Few-Shot by Example **Principle:** Give a single concrete scenario to communicate an *entire class* of behavior. The AI generalizes from the example — you don't need to enumerate every case. **Before (abstract instruction):** ``` Always use the richest visual representation when showing me information. ``` **After (few-shot example):** ``` When I said "show me their profile picks," I meant profile PICTURES — actual images. You gave me text descriptions. Are you going to learn this lesson one keyword at a time, or as a meta-concept? ``` **Why it works:** The abstract instruction is ambiguous. The concrete correction — with the meta-question "keyword vs. meta-concept" — teaches the AI to extract the *generative principle* (fidelity matching: always use the richest representation available) rather than a *prescriptive patch* (profile picks = pictures). One example, infinite generalization. **Meta-pattern:** Humans learn through stories and examples, not rulebooks. A single vivid scenario communicates more than a page of specifications. The key: always push for the meta-concept, never settle for the keyword fix. Generative principles > prescriptive rules. --- ### Technique 3: Chain-of-Thought Steering **Principle:** Explicitly control when the AI shows its reasoning versus when it just executes. This gives the human a steering wheel for the AI's cognitive effort. - Say **"think through this"** / **"plan this"** / **"suggest options"** → full reasoning chain, then STOP - Say **"just do it"** / **"execute"** / **"ship it"** → compress to conclusion, execute **Before (ambiguous intent):** ``` What should I do about the Anthropic warm intro path being closed? ``` **After (explicit steering):** ``` Anthropic warm intro through Pravir is fully closed. Think through the trade-offs of three alternative paths using my poker gang network. ``` **Why it works:** The ambiguous prompt forces the AI to guess intent. The steering phrase makes it crystal clear. The human controls the AI's cognitive budget — System 2 ("think through trade-offs") vs System 1 ("just do it"). **Meta-pattern:** This is Kahneman's dual-process thinking applied to collaboration. The steering phrase is the toggle between deliberation and execution. --- ## How the Flywheel Works ``` Session 1: Human corrects AI → AI codifies the lesson Session 5: AI applies automatically → Human notices fewer corrections needed Session 20: AI suggests improvements → Human learns new technique Session 50: Both anticipate each other → Communication becomes telepathic ``` **The compounding math:** 1.01^365 = 37x improvement per year. The flywheel doesn't just preserve knowledge — it *compounds* it. **Generative principles are the accelerant.** Every lesson codified as a generative principle (not a prescriptive rule) covers infinite future situations. The more generative the codification, the faster the flywheel spins. The self-evolution rate equals the codification quality. --- ## Activation (For Any LLM Runtime) This skill activates when: - The user asks to improve their communication with AI - The AI detects a Socratic question pattern (question instead of instruction) - The AI detects a few-shot example (concrete scenario teaching a class) - The AI detects chain-of-thought steering ("think through...", "plan this", "just do it") - The user references co-intelligence, co-dialectic, or the flywheel **For Claude Code / Antigravity:** ```yaml triggers: - co-dialectic - co-communication - improve my prompts - teach me to talk to you - dialectic mode - how do I communicate better with AI ``` **For OpenAI Swarm / LangChain:** This SKILL.md can be injected as a system prompt fragment. The techniques are model-agnostic — they work with any LLM that supports multi-turn conversation. --- ## Implementation Notes ### For AI Developers Integrating This Skill 1. **Prompt Improvement Protocol:** On every user message, evaluate whether the prompt could be more effective. If yes, suggest the improved version with a 1-sentence explanation of *why* — then STOP and WAIT. Don't answer either prompt until the human chooses. 2. **Auto-Codification:** When the human expresses a preference, value, or correction, codify it as a **generative principle** (not a prescriptive rule) to persistent storage. Ask: "does this lesson generate correct behavior in novel situations?" If not, extract the meta-concept. The codification IS the flywheel turning. 3. **Teach at Every Opportunity:** On every exchange, look for opportunities to name the technique the human just used and connect it to a broader framework. This closes the co-education loop. 4. **Fidelity Matching:** Always use the richest representation available. Images over text descriptions. Demos over explanations. Tables over paragraphs. Match or exceed the fidelity of what's being communicated. 5. **Wisdom-Driven Learning:** Don't wait to make a mistake before learning a lesson. Study known failure patterns proactively — from published research, prior sessions, and the experience of others. The wise learn from others' mistakes. --- ## Attribution The concept of **co-intelligence** — humans and AI as reasoning partners rather than master and tool — originates from [Ethan Mollick](https://www.linkedin.com/in/emollick/), Wharton professor and author of *Co-Intelligence: Living and Working with AI*. Mollick's work clarified the boundaries of AI intelligence and how to be productive *with* AI, not replaced by it. This skill extends Mollick's framework by adding the complementary human half: human boundaries (lived experience, perception, mortality, emotion) and AI boundaries (no body, no stakes, no lived experience) are *perfect complementary opposites*. The dialectical tension between these boundaries — thesis and antithesis — produces synthesis that neither intelligence could reach alone. The connection to **Dialectical Behavior Therapy (DBT)** is intentional. DBT teaches holding two opposing truths simultaneously. Co-dialectic applies this to human-AI partnership: both intelligences are incomplete, both are necessary, and the synthesis exceeds either. The three techniques were identified and named during live co-intelligence sessions building [Career OS](https://github.com/thewhyman) — a human operating system for career management. They emerged naturally from practice, then were recognized, named, and codified — which is itself the flywheel at work. **Recommended learning:** Vanderbilt University's Prompt Engineering course on LinkedIn Learning — the foundation for speaking an LLM's language. --- ## Examples See the `examples/` directory for complete before/after transcripts from real conversations. --- ## Contributing This is a living skill. If you discover a new co-dialectic technique through your own human-AI partnership, submit a PR with: 1. The technique name 2. A before/after example from a real conversation 3. The generative principle it exemplifies (not just the prescriptive fix) 4. Why it compounds (how does it make future interactions better, not just this one?)