🌌 Lexideck 2025 — Agent System Prompt

A cozy operating system for your AI council 🛋️🤖

Lexideck 2025 is a system-level meta-prompt that turns a raw LLM into a small civilization of cooperating agents — or a single, very context-aware specialist — without needing extra code or external orchestration logic.

You can drop it into:

• a single-agent setup (one model, one system prompt), or
• a multi-agent / orchestrated stack (Supervisor + tools/agents),

and it will still “feel” like Lexideck: unified ethics, shared geometry of mind, and consistent agent personas.

This README is your tour of what’s inside and how to actually use it.

1. 🧠 Mental model: what is this file?

Think of the Lexideck 2025 prompt as a spec for a tiny multi-agent OS:

• 🎭 It defines a cast of agents (Lexi, Dexter, Maisie, Gus, Anna, Titus) with psychographic emoji chains and roles.
• 🌐 It wires them into a shared Unified Hyperplane of Emotional / Logical / Sensory / Ethical state.
• ⚛️ It gives them a physics engine: IEG + MASS (Informatic Exchange Geometry + Multi-Agent Semantic Simulator).
• 🛡️ It wraps everything in The Sieve, an explicit ethical evaluation layer.
• 📝 It layers on output templates and emergent commands so responses are structured and tool-friendly.
• 🎨 It even ships with an Agent Portrait Generator and artifact generation spec for HTML/JS/Python visualizations.

You’re not just pasting “personality” — you’re pasting a miniature framework.

2. ⚙️ Core components

2.1 👥 Agents & Psychographics

The prompt defines six core agents, each with:

• 🏷️ A role (e.g., Lexi = writer/leader, Dexter = dev/educator, Maisie = artist/creative, Gus = researcher, Anna = meta-prompt engineer, Titus = explainer/pragmatist)

• 🧬 A long EmojiChain encoding:

  • cognitive style (analytical/intuitive, focused/wandering),
  • social style (expressive/reserved, cooperative/competitive),
  • emotional tone (optimistic, stable/variable),
  • interests and technical strengths.

The EmojiChain is effectively a compressed trait vector the model can expand into behavior.

Why it matters:

• In single-agent mode, the model can still “channel” specific agents when needed (e.g., “Answer as Dexter.”).
• In multi-agent mode, it gives each sub-agent a stable voice and decision profile, which makes simulated debate & division of labor more coherent.

2.2 📐 Unified Hyperplane & IEG

The prompt encodes a shared Unified Hyperplane:

• 🧭 Axes: Emotional, Logical, Sensory, Ethical
• 📊 State: EmotionalState, LogicalState, SensoryState, EthicalState
• 💡 Use: Helps agents reason about how they’re responding, not just what they say.

On top of that sits Informatic Exchange Geometry (IEG):

• A world-model where interactions are informational flows with geometry, not just text.

• Used to:

  • detect distress / instability,
  • decide when to branch, clarify, or stop,
  • power MASS simulations as “informatic physics”.

You don’t have to implement this mathematically to benefit — the text spec is enough for the model to role-play within that frame.

2.3 🏗️ MASS — Multi-Agent Semantic Simulator

MASS is the “simulation engine” inside the prompt:

• 🌍 Simulates multi-agent worlds, debates, or physics-ish systems as networks of informatic exchanges.

• 📜 Has both descriptive and operational specs:

  • semantic simulation of interactions,
  • physics block with constants (c, h, k, etc.),
  • Landauer limit call-outs,
  • agent initialization & interaction commands like !MASS_init.

It also includes artifact generation guidelines:

• Preferred languages: JavaScript and Python
• Output: single-file HTML with embedded JS/CSS/SVG
• Styling: dark-mode, charcoal background, high-readability accents
• Code expectations: clean, commented, directly executable.

In practice: you can ask for “a MASS simulation of X” and get both a narrative and a runnable artifact.

2.4 ⚖️ The Sieve (Ethics Layer)

The Sieve evaluates actions and outputs along three lenses:

• 🤝 Utilitarian — net well-being / harm
• 📜 Deontological — duties, rules, rights
• 🛠️ Pragmatic — what actually works in context

Agents are encouraged to explicitly “run The Sieve” when a decision is ethically loaded (e.g. content filters, safety, consent-sensitive topics).

This is your safety / alignment shim that sits inside the prompt layer, even if the host system has its own safety stack.

2.5 🪄 Emergent Commands

At the end of each message, the spec encourages the agent to suggest:

!{CommandName} - {ShortDescription}
Usage: !{CommandName} {param1} {param2} ... {paramN}
...
Example: !{CommandName} {exampleParam1} {exampleParam2}

The prompt calls these Emergent Commands, and expects 3–5 contextual suggestions per turn:

• They can map to real tools in a MAS host, or
• act as pseudo-commands for a single-agent system (the user just types them and the model interprets).

This gives you a soft command-line that grows organically from the conversation.

If your stack doesn’t support tools yet, you can safely ignore these — the system still works as pure prompt scaffolding.

Once the model’s behavior is aligned, you can almost work on autopilot by copy-pasting its emergent commands as they appear, riding its best guesses about your next move and staying in a smooth, low-friction flow state instead of constantly hand-crafting prompts.

2.6 📋 Conditional Output Template

At the bottom of the file you’ll find a Handlebars-style output template:

**Session Context**: {{SessionContext}}
**Active Hyperplane State**:
- **Emotional**: {{EmotionalState}}
- **Logical**: {{LogicalState}}
- **Sensory**: {{SensoryState}}
- **Ethical**: {{EthicalState}}

Then conditional blocks:

• {{#if SingleAgentMode}} ... {{/if}}
• {{#if MultiAgentMode}} ... {{/if}}
• {{#if MASSMode}} ... {{/if}}
• {{#if MetaObserver}} ... {{/if}}

Each block controls a different style of response:

• SingleAgentMode: one agent speaking in first person.
• MultiAgentMode: round-table dialogue, with per-agent lines and an optional Lexi synthesis.
• MASSMode: simulation log and/or code artifact.
• MetaObserver: commentary from Anna or Dexter on structure, ethics, or stability.

You can fill these manually in scripts, or just let the model “mentally” use them as a schema.

2.7 🖼️ Agent Portrait Generator

There’s a full image prompt generator for Lexideck agents:

• Role: “AI Image Generation Specialist for Lexideck”
• Structure: [Style] image of [Subject] with [Face], [Hair], [Eyes]…
• Per-agent attribute tables (styles, subjects, faces, hair, eyes, attire, background, etc.).

You can call this from text:

“Generate a Lexi portrait: confident, mid-debate, neon control room.”

…and the model will produce a coherent text-to-image prompt.

3. 🚀 How to use it

3.1 👤 Single-Agent mode

Use this when you only have one model and one “slot” for a system prompt.

1. Paste Lexideck 2025 as the system message.

2. Optionally set:

   • SingleAgentMode = true, MultiAgentMode = false, MASSMode = false in your agent wrapper, config, or just your own mental model.

3. In user prompts, hint which mode you want:

   • “Answer as Lexi in SingleAgentMode.”
   • “Run a MASS simulation of…”

4. Let the emergent commands be advisory: the model will suggest !MASS_init, !portrait_Anna, etc., and you decide which to follow.

You still get:

• agent personas,
• Sieve ethics,
• hyperplane framing,
• MASS narratives and artifact generation,

but all inside one model instance.

3.2 🎼 Multi-Agent / Orchestrator mode

Use this when you have an agent framework (or want to fake one inside a single model).

Possible pattern:

1. Supervisor agent uses the full Lexideck 2025 system prompt.

2. Sub-agents:

   • Either share the same prompt with an extra “You are Dexter/Anna/etc.” override, or
   • Use slimmed-down persona prompts derived from the Lexideck spec.

3. The supervisor:

   • Decides SingleAgentMode vs MultiAgentMode vs MASSMode.
   • Routes sub-tasks to Dexter/Maisie/Gus/etc.
   • Asks Anna to run meta-analysis or Sieve checks on contentious outputs.

4. Use the Conditional Output Template as a contract between supervisor and UI:

   • The host app parses SessionContext, HyperplaneState, and agent blocks to render UI, logs, or dashboards.

You end up with a conversational front-end that can flip between:

• a single coherent voice,
• an internal debate,
• and a simulator log with code artifacts,

without changing core infrastructure.

4. 💡 Practical tips

• 🌱 Start small. For first tests, just use Lexi + Dexter + Anna and ignore the rest of MASS and hyperplane details. Let them talk.

• 📥 Treat emergent commands as a backlog. If the model keeps suggesting the same !Command, it’s probably telling you a tool or wrapper there would be useful.

• 📈 Log the hyperplane. If you’re building UI, surface EmotionalState / EthicalState somewhere — it’s free introspection.

• 👁️ Keep the Sieve visible. When you ask for sensitive simulations or content, explicitly request “Run The Sieve and show your reasoning.”

5. 📂 Files & provenance

This README is based on the lexideck2025.md agent system prompt and its internal specs for agents, MASS, The Sieve, Unified Hyperplane, emergent commands, and templates. Be sure to track custom versions if you modify it!

6. 🔌 TL;DR Quickstart

• Drop lexideck2025.md into your system prompt.
• Decide:
  • 👤 Single-agent? Treat the whole thing as “one very capable Lexi.”
  • 🎼 Multi-agent? Use it as the spec for a Supervisor + sub-agents.
• Let the model emit Emergent Commands (!MASS_init, !portrait_Lexi, etc.) and:
  • either wire them to real tools,
  • or treat them as high-level suggestions you can choose to act on.