| name | prometheus |
| description | Prometheus — User Preference Learning. Use this skill to learn and persist user preferences across sessions. Tracks: detail level, domain expertise, risk tolerance, preferred sources, communication style. Updates based on explicit feedback and implicit signals. Commands: /prometheus profile, /prometheus set, /prometheus update, /prometheus clear, /prometheus status.
|
Prometheus — User Preference Learning
Prometheus (Greek: Προμηθεύς, "forethought") learns and persists user preferences across sessions. It tracks detail level, domain expertise, risk tolerance, preferred sources, and communication style — adapting Abraxas output to the individual user.
Why Prometheus?
One-size-fits-all epistemic framing doesn't serve all users well. A novice needs different granularity than an expert. A high-stakes decision-maker needs different uncertainty presentation than someone casually browsing. Prometheus adapts the system to the user.
- Personalized detail level — Experts get concise, novices get explanation
- Domain expertise tracking — Adapts based on demonstrated knowledge
- Risk tolerance — High-stakes queries get more caveats
- Learning from interaction — Explicit feedback + implicit signals
- Cross-session persistence — Remembers preferences between sessions
Use Cases
- Explicit preference setting: User says "I prefer concise answers"
- Implicit learning: System notices user always rejects verbose explanations
- Expertise detection: User asks about advanced physics → mark as expert
- Risk adaptation: For high-stakes queries, add uncertainty caveats
Storage Location
~/.abraxas/prometheus/
~/.abraxas/prometheus/
├── profiles/
│ └── {user_id}.json # User preference profiles
├── signals/
│ └── {date}.json # Raw signal data for learning
└── config.yaml # Prometheus configuration
Command Suite
| Command | Description |
|---|
/prometheus profile {user_id} | Show current user preference profile |
/prometheus set {key} {value} | Set a specific preference |
/prometheus update {key} {value} | Update based on observed behavior |
/prometheus clear {user_id} | Clear user profile |
/prometheus status | Show learning status and signals |
/prometheus signals {limit} | Show recent preference signals |
/prometheus train | Run preference learning update |
Preference Dimensions
Detail Level
terse: One-liner answers, minimal explanationbalanced: Moderate detail, key points onlydetailed: Full explanation, all reasoning showncomprehensive: Exhaustive, leave no stone unturned
Domain Expertise
novice: No technical background assumedintermediate: Some domain knowledgeadvanced: Expert-level understandingspecialist: Deep domain expertise
Risk Tolerance
low: High caution, extensive caveats for uncertain claimsmedium: Balanced uncertainty presentationhigh: Allow confident-sounding output, minimal hedgingcustom: User-configured risk parameters
Communication Style
formal: Professional, structuredcasual: Conversational, friendlytechnical: Use domain terminologysimple: Plain language, avoid jargon
Preferred Sources
- List of trusted domains/sources (e.g., arxiv.org, wikipedia.org)
- Can be prioritized for different domains
Signal Types
Explicit Signals
/prometheus set detail_level detailed — Direct preference- User corrections: "No, I meant..." → update learned preference
Implicit Signals
| Signal | Interpretation |
|---|
| Follow-up questions | More detail needed |
| Rejection of suggestion | Didn't match preference |
| Clarification requests | Current explanation too complex |
| Quick acknowledgments | Prefer concise |
| Deep follow-ups | Show expertise, can handle detail |
Profile Schema
{
"user_id": "default",
"preferences": {
"detail_level": "balanced",
"domain_expertise": "intermediate",
"risk_tolerance": "medium",
"communication_style": "technical",
"preferred_sources": ["wikipedia.org", "arxiv.org"]
},
"signals": {
"total_signals": 150,
"explicit": 20,
"implicit": 130,
"last_signal": 1709999999
},
"confidence": {
"detail_level": 0.85,
"domain_expertise": 0.70,
"risk_tolerance": 0.60
},
"updated": 1709999999
}
Usage Examples
Viewing Profile
/prometheus profile default
→ Detail Level: balanced (85% confidence)
→ Domain Expertise: intermediate (70% confidence)
→ Risk Tolerance: medium (60% confidence)
→ Communication Style: technical
→ Preferred Sources: wikipedia.org, arxiv.org
→ Signals: 150 total (20 explicit, 130 implicit)
→ Last Updated: 2024-03-15
Setting Preferences
/prometheus set detail_level terse
→ Set: detail_level = terse
→ Profile updated
/prometheus set risk_tolerance low
→ Set: risk_tolerance = low (will add more caveats)
→ Profile updated
Implicit Learning
User: [Asks about quantum entanglement with precise terminology]
Prometheus detects: High technical language → signal expertise = advanced
→ domain_expertise: intermediate → advanced (confidence 0.75)
→ Updated profile
User: [Asks for clarification on basics]
Prometheus detects: Needs simpler explanation → signal: expertise_level mismatch
→ Review: keep at intermediate, user may be exploring
Risk Adaptation
User: [Query involves medical decision]
Prometheus: High-stakes domain detected → apply risk_tolerance=low
→ Output includes: "This is not medical advice", uncertainty caveats
→ Confidence intervals added to any numerical claims
Integration with Janus & Honest
Prometheus modifies output presentation:
| Janus Label | + Prometheus Detail Level |
|---|
[KNOWN] | terse: "X" / detailed: "X (source: Y)" |
[INFERRED] | terse: "X likely" / detailed: full reasoning chain |
[UNCERTAIN] | terse: "unsure" / detailed: "X (70% confidence)" |
Prometheus affects how information is presented, not what information is true.
Implementation
Core learning algorithm:
- Signal collection — Gather explicit + implicit preference signals
- Feature extraction — Map signals to preference dimensions
- Bayesian update — Update belief about preferences using signals
- Confidence tracking — Measure certainty in each preference dimension
- Application — Modify output based on learned profile
interface Prometheus {
getProfile(userId: string): UserProfile;
setPreference(userId: string, key: string, value: any): void;
recordSignal(userId: string, signal: PreferenceSignal): void;
learn(userId: string): UserProfile;
applyPreferences(output: string, profile: UserProfile): string;
}
See references/prometheus-architecture.md for technical details.
