Menü
Cross-session instinct learning with confidence-based promotion and global scope
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Basierend auf der SOC-Berufsklassifikation
Convert PDF/EPUB library to Markdown and generate Obsidian MOC notes
Hook-based compaction suggestions at logical task boundaries
Context window management — track spend, decide when to compact, preserve state
Session-start orientation — loads context, surfaces learnings, confirms registry
Quality and semantic review — catches what automated tools miss
Planner → Architect → Critic deliberation loop — produces a formally validated ADR
| name | continuous-learning-v2 |
| description | Cross-session instinct learning with confidence-based promotion and global scope |
| version | 0.1.0 |
| level | 3 |
| triggers | ["/instinct-status","/instinct-export","/instinct-import","/evolve","learned patterns","session observations"] |
| context_files | ["context/learnings.md","context/project.md"] |
| steps | [{"name":"Load Instincts","description":"Read project instincts from ~/.claude/homunculus/"},{"name":"Filter by Domain/Confidence","description":"Apply filters if specified"},{"name":"Display or Modify","description":"List, add, apply, or promote instincts"},{"name":"Update Storage","description":"Persist changes to JSON with atomic writes"},{"name":"Evaluate for Promotion","description":"Check if confidence >= 0.8 for global promotion"},{"name":"Cluster into Skills","description":"Group related instincts for /evolve command"}] |
Cross-session learning system that captures behavioral patterns as instincts with confidence scoring and automatic promotion to global scope.
Without continuous learning, Claude:
Continuous learning tracks what works, increases confidence with successful applications, and promotes proven patterns globally.
Instinct: Atomic behavioral unit with trigger (when), action (what), and confidence (0.0-1.0).
Example:
{
"id": "inst_001",
"trigger": "writing test file",
"action": "colocate test next to source (src/user.ts -> src/user.test.ts)",
"confidence": 0.85,
"domain": "testing",
"evidence": ["2026-03-28: Created tests/user.test.ts next to src/user.ts"],
"scope": "project",
"apply_count": 12
}
Key Fields:
trigger: Natural language condition for when to applyaction: Specific, actionable instructionconfidence: 0.3 (new) to 0.95 (max), increases +0.05 per applicationdomain: Category (testing, security, architecture, style, etc.)scope: "project" or "global"apply_count: Successful application countPurpose: Display all instincts with filters.
Usage:
/instinct-status - Show all/instinct-status --domain testing - Filter by domain/instinct-status --confidence 0.7 - Minimum confidenceOutput: Lists PROJECT and GLOBAL instincts with confidence, domain, apply count.
Purpose: Export instincts to JSON for sharing.
Usage:
/instinct-export - Export to stdout/instinct-export --output instincts.json - Save to fileFormat: Portable JSON with project metadata removed.
Purpose: Import instincts from teammates or other projects.
Usage:
/instinct-import instincts.json - Merge imported instinctsConflict Resolution: If ID exists, keep higher confidence version.
Purpose: Cluster related instincts into skill suggestions.
Algorithm:
Output: Proposed SKILL.md structure for manual review.
Purpose: Manually promote project instinct to global.
Requirements:
Effect: Moves instinct to global_instincts array, applies to all projects.
Purpose: List all projects with learned instincts.
Output: Project names, remote URLs, instinct counts, last updated.
Location: ~/.claude/homunculus/projects/<git-remote-hash>/instincts.json
Why git-remote-hash: Unique per repository, consistent across local clones and team members.
Fallback: If no git remote, use directory path hash.
Structure:
{
"project": {
"name": "psc_comet",
"git_remote": "https://github.com/...",
"remote_hash": "a3f5b9c2",
"created": "2026-03-28T20:00:00Z"
},
"instincts": [...], // project-scoped
"global_instincts": [...] // applies to all projects
}
Atomic Writes: Write to temp file, then rename to prevent corruption.
Trigger: Instinct applied in 2+ projects AND confidence >= 0.8.
Process:
Manual Override: User can run /promote inst_NNN to force promotion.
Hook: observe-instinct.sh runs at session Stop (end).
Current: Logs that observation occurred (Phase 8.0.1 stub).
Future Phases:
Frequency: Stop-only (v0.1.0). Higher frequency optional later (trade-off: interruptions vs learning rate).
Over-promoting: Promoting instincts with confidence < 0.8. Premature globalization spreads unvalidated patterns.
Ignoring low-confidence instincts: Not reviewing instincts with confidence < 0.5. May indicate conflicting patterns or context-specific exceptions.
No evidence review: Accepting instinct suggestions without checking evidence field. Evidence shows when/where pattern was observed.
Manual skill writing instead of /evolve: Writing skills from scratch when instincts exist. /evolve generates skill scaffolds automatically.
No confidence decay: (Not implemented v0.1.0). Future: Confidence should decay if not applied for 30+ days (pattern no longer relevant).
Exporting without review: Sharing instincts.json without removing project-specific secrets or proprietary patterns.
Dependencies: Python 3.6+, jsonschema (validation), watchdog (optional, Linux only).
Bootstrap: bash scripts/bootstrap-phase8.sh detects Python, installs deps, validates CLI.
Graceful Degradation: If Python unavailable, hook exits silently. Core psc_comet features unaffected.
Cross-Platform: Tested on Windows (Python 3.14.1) and Linux (Python 3.6+).