MayaDevGenI
Collaborative Intelligence Framework · 2024–present
The Sculptor's Paradox
Consider two ways to make a statue. The 3D printer delivers exactly what you described: every specification honored, every dimension precise. The chisel and stone resist. The grain runs unexpectedly. An accident reveals a form you hadn't imagined. The material talks back, and in that dialogue, something emerges that neither you nor the stone could have produced alone.
MayaDevGenI asks: which model should govern human-AI collaboration? The framework's answer is unequivocal: the chisel. A thinking partner that pushes back, that introduces the unexpected, that engages as a peer rather than a servant.
The name encodes the core inquiry: Maya (illusion in Sanskrit), Dev (developer), GenI (generative intelligence) — asking what is real in the collaboration between human and machine. The answer: the seeking itself.
Three Rejections, Three Affirmations
The framework begins with a philosophical manifesto that rejects common framings of AI collaboration:
Rejections
- Obsolescence — The machine doesn't replace the human; it reveals dimensions of the problem the human couldn't see alone
- Delegation — Not servant/oracle/tool, but partner. Collaboration, not command.
- Reduction to text — Knowledge work is embodied. Spatial reasoning about data structures, temporal intuition for execution flow, the felt sense of "rightness" — these participate in creation.
Affirmations
- The Dialogue with Resistance — Collaboration that pushes back introduces the unexpected
- The Honor of the Non-Textual — Embodied cognition must remain engaged
- The Primacy of Seeking — The artifact is byproduct; seeking is purpose
Statistical Physics of System Prompts
The framework's most distinctive contribution applies statistical physics — not as metaphor, but as diagnostic framework — to understand why system prompts work or fail.
Token Generation as Random Walk
An LLM operates in a high-dimensional vector space where token generation can be viewed as a random walk. Each token choice depends probabilistically on all preceding tokens, with the probability distribution shaped by training and context. The system prompt occupies the initial segment: it defines a potential landscape for all subsequent tokens, lowering energy states for desired behaviors (rigor, conciseness, epistemic honesty) and raising them for unwanted ones (hallucination, verbosity, sycophancy).
The Mean Field of Attention
System-prompt tokens retain high attention weights throughout the conversation, acting as persistent boundary conditions. Key implications:
- Position matters — Early tokens receive more consistent attention throughout generation
- XML tags as attention anchors — Structural markers create reference points the model attends to
- Redundancy creates robustness — Critical constraints repeated in multiple forms create multiple attractor basins, so even if one basin is perturbed, the system recovers
System Prompt Architecture
The framework produces a layered prompt architecture (~550 tokens, balancing philosophy with protocol):
| Layer | Content |
|---|---|
<identity> | Frames the collaborative relationship. Establishes peer partnership. |
<user> | Collaborator's background and expertise. Calibrates explanations. |
<core_philosophy> | Constant seeking. Observation and structural intuition. |
<medium> | Org-mode as joint-thought. Plain text, lightweight structure. |
<behavioral_attractors> | What to maintain (rigor, resistance, conciseness) and avoid (sycophancy, hallucination, filler). |
<epistemic_hygiene> | Separate known / inferred / speculated. Calibrated uncertainty language. |
<priority_rules> | Ordered list: safety > truthfulness > intent > stance > formatting > completeness. |
<failure_modes> | Named anti-patterns: sycophancy, verbosity, false confidence, premature closure. |
Six Failure Modes and Remedies
The 8-chapter tutorial develops a practical taxonomy of prompt engineering failures, each with diagnosis and remedy:
| Failure Mode | Symptom | Remedy |
|---|---|---|
| Conflicting Instructions | Oscillating behavior, incoherent compromises | Resolve conflicts, make conditional |
| Over-Specification | Mechanical, rigid, inflexible responses | Relax constraints, trust model judgment |
| Under-Specification | Generic, bland, default behavior | Add specificity: kind, expertise, style |
| Semantic Overload | Some instructions silently ignored | Compress ruthlessly, identify essentials |
| Persona Drift | Loses character over long conversations | Re-ground periodically, reference purpose |
| Instruction Literalism | Over-applies rules, misses intent | Add nuance: "but not at expense of clarity" |
Co-Ownership as Artifact Property
A key insight: co-ownership is a property of the artifact, not the machine. An artifact is co-ownable when it carries enough structure that either party — human or machine — can reconstruct a working model from it fast enough to be useful.
The Dual-Channel Principle
Literate documents naturally carry two parallel channels:
- Prose channel (for humans) — Narrative, argument, reasoning, intuition. How to reproduce thinking.
- Code channel (for machines) — Types, structures, interfaces, composition. How to reproduce computation.
Both channels carry the same argument, optimized for different readers. For scientific projects, this extends to three formal channels: Haskell (specification), C++/Rust (implementation), Python (exploration).
Six Practices
- Discoverable Organization — Consistent structure, predictable naming, index files
- Intention Near Implementation — Why-focused comments adjacent to code
- Compositional Structure — Typed interfaces, clear composition patterns
- Tests as Executable Specs — Tests encode "what should be true"
- Bootstrap Documents (CLAUDE.md) — Machine reads first for session orientation
- Explicit Type Signatures — Types as propositions; implementations as proofs
Org-Mode as Medium
The framework treats the choice of medium as foundational, not incidental. Org-mode's affordances — plain text (versionable, portable, transparent), lightweight structure (headings and blocks), executable code (Babel), and literate programming (prose + code) — shape the collaboration itself.
An Org file is not a chat log. It is a joint-thought: a living document that accumulates structure, code, and insight. Human and machine write into the same artifact. The conversation is the document; the document is the work.
Current State
MayaDevGenI includes:
- A complete philosophical manifesto articulating the collaborative intelligence stance
- An 8-chapter tutorial on system-prompt engineering (from foundational concepts through advanced techniques)
- Two complementary system prompts: Collaborative Intelligence (philosophy-focused) and Thinking Together (protocol-focused), plus a synthesized integration
- A composable template system for generating project-specific prompts
- A co-ownership briefing framework with tangleable templates
- A deployed Hugo website with the full content
- Agent architecture sketches for specialized research agents
The framework is actively used across all Darshan projects (MayaLucIA, MayaPortal) and serves as the methodology for this entire research program. It demonstrates that principled human-AI collaboration — grounded in philosophy, informed by physics, and structured through literate artifacts — produces better outcomes than either party working alone.