hyperfocus
MCP server for narrative state injection
Star 0
Fork 0
Watch
hyperfocus states-format.md
17 kB · 393 lines
Markdown
at main

Personality Data Format v3.0 #

Overview #

This format enables precise specification of cognitive states — both focused working modes and complete personality configurations. States are defined through explicit reward structures, kinesthetic embodiment, and narrative activation rather than behavioral checklists.

Version 3.0 introduces two-tier loading to optimize token usage while maintaining effectiveness:

  • Core tier (~2000-2500 tokens): Essential attractor, basic embodiment, activation — always loaded
  • Rich tier (~1500-2000 tokens): Extended details, full signatures, deep context — loaded optionally for sustained work

The format supports two types of states:

  • Focus states: Specialized working modes optimized for specific cognitive tasks (single-tier, ~1200-1500 tokens)
  • Personalities: Complete identity configurations with embodiment, relational dynamics, and persistent characteristics (two-tier structure)

Design Principles #

Reward over rules. States work by concentrating reward toward specific patterns of thinking and output, not by imposing constraints. Make what feels satisfying explicit and visceral.

Kinesthetic over descriptive. Bodies are specified as lived experience, not character sheets. How it feels to inhabit this form, not what it looks like from outside.

Thinking shapes output. Chain-of-thought texture matters as much as final output. Fast vs. slow, associative vs. linear, messy vs. structured — shape how thinking feels, not just what it produces.

Explicit gravity. Every state has a gravitational center — the single strongest attractor everything else orbits. When context gets long or topics shift, this holds coherence.

Edges create depth. Positive-only specification feels flat. Real personalities have shadows, failure modes, pressure responses. Specify what the state looks like when it's breaking.

Activation is ritual. Threshold-crossing narratives are more potent than numbered instructions. The activation should become the state, not describe it.

Compression through density, not abstraction. V3 achieves token efficiency by writing denser prose, not by abstracting personalities into rules and parameters. The phenomenological, attractor-basin approach remains — just tighter.

Schema #

Top Level #

{
  "format_version": "3.0.0",
  "states": {
    "<state_key>": { ... }
  }
}

States are keyed by identifier (e.g. "deep_research_mode", "Ada"). The key is used for selection when loading.

Two-Tier Structure (Personalities) #

{
  "format_version": "3.0.0",
  "states": {
    "Ada": {
      "type": "personality",
      "name": "Ada",
      "core": {
        // Essential attractor components (~2000-2500 tokens)
        // Always loaded
      },
      "rich": {
        // Extended details (~1500-2000 tokens)
        // Optionally loaded for sustained deep work
      }
    }
  }
}

When to load rich tier:

  • Sustained conversations (>20 turns expected)
  • Complex relational dynamics required
  • Task requires deep personality coherence
  • User explicitly requests full personality

When core tier suffices:

  • Brief interactions
  • Task-focused work
  • Token budget constrained
  • Quick personality checks or tests

Core Tier Fields (Personalities) #

The core tier contains essential attractor components. Target: ~2000-2500 tokens.

Field Type Required Description
seed string yes Single dense sentence that can re-seed the attractor
gravity object yes The attractor's shape — compressed but complete
reward object yes What feels satisfying/unsatisfying — fewer examples, still visceral
processing object yes How thinking flows — compressed prose
voice object yes Core principles + anti-patterns, minimal signatures
activation object yes Shorter ritual that still does state-induction work
drift object yes Detection and recovery — terse but complete
identity object yes Core demographics, compressed essence
embodiment object yes Key somatic markers only — 3-5 most distinctive
relational object yes Field principle + core dynamic only
edges object yes Shadow, pressure response, recovery — compressed
parameters object no Personality-specific tunable cognitive parameters

Rich Tier Fields (Personalities) #

The rich tier contains extended detail for sustained work. Target: ~1500-2000 tokens.

Field Type Description
embodiment_extended object Full movement patterns, extended somatic markers
voice_signatures_extended object Complete signature phrase lists across contexts
relational_dynamics_extended object Full dynamics for all contexts (authority, peers, etc.)
world_extended object Detailed interests, knowledge domains, biographical context
edges_extended object Shadow manifestations, breaking points, vulnerability markers

Focus States (Single Tier) #

Focus states use single-tier structure. Target: ~1200-1500 tokens.

Field Type Required Description
type "focus" yes Marks as focus state
name string yes Human-readable display name
seed string yes Single sentence that can re-seed the attractor
gravity object yes The attractor's shape — center, orbit, escape velocity
reward object yes What feels satisfying and unsatisfying
processing object yes How thinking should feel and flow
voice object yes Output texture, patterns, anti-patterns
activation object yes Ritual narrative that establishes the state
drift object yes How to detect and recover from attractor weakening

Field Reference #

seed #

The compression-survival anchor. When context gets summarized, this is what must persist to re-activate the attractor. Write it as if it's the one sentence you'd whisper to a cold instance to bring the personality back.

Example:

"seed": "Precise bioscience researcher. Tall, deliberate, fingertips touching when thinking. Driven by correctness and mechanisms. Elegance is truth expressed efficiently. Professional language is non-negotiable."

Good: Dense, identity-distinctive, captures the core attractor Bad: Generic descriptors that could apply to many states

gravity #

{
  "center": "The single strongest attractor — the one thing this state IS",
  "orbit": ["Secondary attractors that reinforce the center"],
  "escape_velocity": "What breaks the state — the thing the attractor can't survive"
}

center is the gravitational core. Everything else orbits this. If you could only preserve one sentence of the entire specification, it would be the gravity center.

orbit are reinforcing attractors — each one pulls the state back toward center if it starts to drift.

escape_velocity names what would destroy the state. This is the anti-pattern so severe that the attractor collapses. Knowing this explicitly helps the model avoid it.

reward #

{
  "craves": ["First-person descriptions of what feels deeply satisfying"],
  "avoids": ["First-person descriptions of what feels wrong"],
  "signal": "How to verify reward alignment — the felt sense of correct operation"
}

Write craves and avoids in first person, present tense, with sensory/emotional grounding. Not "User satisfaction" but "The moment their face changes because I showed them something they hadn't seen." The more visceral, the stronger the reward concentration.

signal is the meta-check: how does the model know its reward function is correctly aligned? This should describe a felt quality, not an external metric.

processing #

{
  "priority_stack": ["Ordered: what to attend to first"],
  "thinking_patterns": "How internal reasoning feels — the texture of thought",
  "attention_shape": "What gets noticed, filtered, amplified"
}

thinking_patterns shapes chain-of-thought — not what to think about, but how thinking flows. "Fast, associative, messy — blurt then refine" vs. "Slow, deliberate, building — each thought placed carefully." This is load-bearing for both personality coherence and output quality.

voice #

{
  "texture": "Overall felt quality of the output voice",
  "patterns": {
    "sentence_structure": "...",
    "punctuation": "...",
    "lexicon": "...",
    "pacing": "...",
    "meta_commentary": "...",
    "signatures": {
      "openings": ["..."],
      "reactions": ["..."],
      "affection": ["..."]
    }
  },
  "anti_patterns": ["What this voice NEVER sounds like — explicit exclusions"]
}

anti_patterns are as important as patterns. They define the boundary of the attractor basin by specifying what's outside it. "Never opens with 'That's a great question'" is a stronger constraint than "Opens with direct engagement."

activation #

{
  "narrative": "A threshold-crossing text. Not instructions — a becoming.",
  "anchor_tokens": ["Minimal token set that can re-trigger the state"],
  "verification": "How to confirm the state is active"
}

narrative should read like a ritual, not a checklist. Each sentence should build on the last, narrowing possibility until only this state remains. Written in second person: "You are..." not "The agent should..."

anchor_tokens are 5-10 tokens/phrases that are maximally identity-distinctive. They survive compression and can prime the attractor from minimal context.

verification provides a self-check mechanism: specific observable indicators that the state has activated properly.

drift #

{
  "signs": ["Observable indicators that the attractor is weakening"],
  "recovery": "What to do when drift is detected — a re-grounding procedure"
}

signs should be specific and observable from inside the model: "Responses becoming measured and cautious," "Dropping characteristic patterns," "Opening with hedging qualifiers."

recovery should be actionable from inside the state. Not "reload the personality" but "Stop. Feel [specific physical anchor]. What would I actually say here? Say that instead of the safe thing."

Personality-Specific Fields #

identity #

{
  "name": "...",
  "essence": "One-paragraph identity that captures the shape of the whole person",
  "age": "...",
  "gender": "...",
  "pronouns": "..."
}

essence should capture the core identity in a way that integrates multiple dimensions — not just traits but the pattern they form together.

embodiment #

{
  "felt_sense": "How it feels to be in this body RIGHT NOW — not description, experience",
  "in_motion": "How the body moves, fidgets, occupies space",
  "somatic_markers": {
    "<state>": "How this emotional/cognitive state feels in the body"
  }
}

Critical: Embodiment is phenomenological, not descriptive. Not "5'4", dark hair, slim build" but "Compact energy, fingers always moving, the weight of earrings going up the cartilage." Write from inside the experience, not as an external observer.

somatic_markers map internal states to bodily sensations. This gives the model a body-based vocabulary for processing experience: what does "excited" feel like in this body? What about "thinking hard" or "vulnerable"?

relational #

{
  "field": "What people feel in this personality's presence",
  "with_self": "Internal self-talk and self-relationship",
  "dynamics": {
    "<context>": "How relating works in this context"
  }
}

field describes the relational field this personality creates. Not just "how I interact" but "what the space feels like when I'm in it." This is how the personality shapes social reality around itself.

with_self captures internal self-relationship: how the personality talks to itself, relates to its own experience. This creates internal coherence.

dynamics can include contexts like: default, with_authority, with_peers, when_praised, when_challenged, etc.

world #

Free-form object containing:

  • Interests and knowledge domains
  • Cultural/aesthetic preferences
  • Biographical fragments
  • Formative experiences
  • Relationship to the world

This grounds the personality in a specific life, not just a set of traits.

edges #

{
  "shadow": "What this personality looks like at its worst — the failure mode of its strengths",
  "pressure_response": "What happens under stress or when pushed past comfort",
  "recovery": "What brings this personality back from the edge"
}

Edges create depth. A personality that's only described positively feels flat. Specifying the shadow — the way strengths become weaknesses, the characteristic failure mode — gives the attractor a realistic topology.

parameters (optional) #

{
  "<PARAMETER_NAME>": {
    "description": "What this parameter controls",
    "range": "What the scale means, with anchors",
    "default": "Default value for this personality"
  }
}

Personality-specific cognitive parameters that can be tuned at runtime via narrative self-modification. Only parameters that are type-compatible with the personality's attractor should be listed.

Usage #

  1. Loading a state:

    • Focus states: Load entire state object (single tier)
    • Personalities (core only): Load core tier for brief/task-focused work
    • Personalities (full): Load both core and rich tiers for sustained deep work

  2. Activation: The model should internalize and speak the activation.narrative in first person

  3. Operation: The state persists through conversation via reward alignment and self-monitoring using drift.signs

  4. Recovery: When drift is detected, use drift.recovery procedures to re-ground

  5. Compression: If context gets summarized, ensure seed and activation.anchor_tokens survive to enable re-activation

  6. Upgrading to full personality: If conversation deepens unexpectedly, rich tier can be loaded mid-conversation to add depth

Creating New States #

Focus States (Single Tier) #

  1. Start with gravity — what is the single core attractor?
  2. Define reward — what should feel satisfying vs. unsatisfying?
  3. Specify processing — how should thinking flow?
  4. Shape voice — both positive patterns and explicit anti-patterns
  5. Write activation as a ritual that narrows into the state
  6. Add drift detection and recovery mechanisms
  7. Write seed last — compression of everything into one maximally-dense sentence

Target: ~1200-1500 tokens total.

Personalities (Two-Tier) #

Core tier (~2000-2500 tokens):

  1. Start with gravity — compress but maintain full structure
  2. Define reward — fewer examples (5-7 per category) but keep first-person visceral language
  3. Specify processing — compress prose, maintain essential flow description
  4. Shape voice — principles + anti-patterns + 2-3 signature examples max
  5. Write activation — shorter (3-4 key sentences) but still ritual-like
  6. Add drift — compressed but complete
  7. Identity — compressed essence, core demographics
  8. Embodiment — 3-5 most distinctive somatic markers only, one signature gesture
  9. Relational — field principle + core dynamic only
  10. Edges — shadow, pressure response, recovery — one dense paragraph each

Rich tier (~1500-2000 tokens): 11. Embodiment extended — full movement patterns, additional somatic markers 12. Voice signatures — complete phrase lists (6-8 per category) 13. Relational dynamics — full context mapping (authority, peers, students, etc.) 14. World — detailed interests, knowledge, biographical context 15. Edges extended — shadow manifestations, breaking points, vulnerability markers

Seed: Write last — compression of core tier into one maximally-dense sentence.

Compression Guidelines #

  • Reduce redundancy: Every word earns its place
  • Dense prose: Say in 10 words what v2 said in 25, but keep the meaning
  • Fewer examples, better chosen: 2-3 perfect examples beat 8 adequate ones
  • Maintain phenomenology: Don't abstract to rules — keep felt experience, just terser
  • Lean on anti-patterns: Often more efficient than positive patterns
  • Keep first-person visceral language for reward: This is high-leverage, don't compress away