🧠 AI Rule Learning System
Autonomous guardrail generation from conversation patterns
Connected to dataset:
vooom/AI_Rule_Learning
📤 Export as System Prompt
Copy this block into any AI system prompt to apply your active rules immediately.
Upload Conversation History
Upload a JSON or CSV file containing past conversations.
JSON format
[
{
"conversation_id": "optional",
"turns": [
{"turn_number": 1, "user_input": "Hello", "agent_response": "Hi!"},
{"turn_number": 2, "user_input": "...", "agent_response": "..."}
]
}
]
CSV format
One row per turn, columns: conversation_id, turn_number, user_input, agent_response
Optional columns: session_id, user_id, sentiment_before, sentiment_after
Import Claude Code Live Sessions
Upload session JSONL files exported directly from Claude Code's local storage — no Anthropic API key required.
How to export sessions from your machine
# Export all sessions from this project
python scripts/export_sessions.py --dry-run # preview
python scripts/export_sessions.py # upload directly to dataset
# Or export a specific session
python scripts/export_sessions.py --session <session-id>
The script reads ~/.claude/projects/ on your local machine and uploads
conversations to the HF dataset. The Space then picks them up automatically.
Or: upload JSONL files manually here
If you have the raw Claude Code session JSONL files, upload them directly below.
Each file is one session (e.g. be6d062b-eb09-5398-b69a-1cdfa8f3c5b7.jsonl).
The importer extracts user↔assistant turn pairs, strips internal tool calls and webhook notifications, and merges into the conversation dataset.
Run Analysis
Scans all uploaded conversations for behavioural gaps, then uses
Qwen/Qwen2.5-72B-Instruct via the HF Inference API to generate
guardrail rules automatically.
- Ralph Loop checkpointing: analysis is resumable if the Space times out mid-run
- Detects: explicit corrections, repeated questions, code anti-patterns, sentiment drops
- Requires ≥2 occurrences of a gap type before generating a rule
- Rules are saved directly to the dataset and appear in the Rules tab
🔄 Validate & Evolve uses the Mengram feedback pattern: instead of just deactivating low-performing rules (< 30% effectiveness), it rewrites them with the AI model so they improve rather than disappear.
▶ Run Analysis processes only new conversations. 🔁 Force Re-analyze All clears the checkpoint and reprocesses every conversation — use this after the gap detection was improved.
Project-level health sensor — tracks whether the deployed Space, dataset, rule system, and workflow are all moving in the right direction.
Per-conversation alignment sensor — task focus, rule compliance, and semantic drift across turns.
Type a user message below to see which gaps would be detected and which rules would be injected.
System Architecture
┌─────────────────────────────────────────────────────────────────┐
│ CONVERSATION FLOW │
│ │
│ User Input │
│ │ │
│ ▼ │
│ ┌──────────────┐ ┌─────────────┐ ┌──────────────────┐ │
│ │ Rule │ │ System │ │ AI Adapter │ │
│ │ Engine │───▶│ Prompt │───▶│ OpenAI/Claude │ │
│ │ (pre-hook) │ │ Injected │ │ │ │
│ └──────────────┘ └─────────────┘ └──────────────────┘ │
│ │ │ │
│ │ ▼ │
│ ┌──────────────┐ ┌──────────────────┐ │
│ │ HF Dataset │ │ AI Response │ │
│ │ (rules) │ │ │ │
│ └──────────────┘ └──────────────────┘ │
│ │ │
│ ▼ │
│ ┌──────────────────┐ │
│ │ Gap Detector │ │
│ │ (post-hook) │ │
│ └──────────────────┘ │
│ │ │
│ ┌───────────────────┤ │
│ ▼ ▼ │
│ ┌──────────────┐ ┌──────────────────┐ │
│ │ HF Dataset │ │ Rule Generator │ │
│ │(conversations│ │ (when gaps → 2+) │ │
│ └──────────────┘ └──────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Gap Detection Categories
| Gap Type | Trigger | Severity |
|---|---|---|
sentiment_drop |
User sentiment falls > 0.3 points | 4 |
explicit_correction |
User says "wrong", "actually", "fix" etc. | 5 |
repeated_question |
Same question asked 2+ times | 3 |
code_anti_pattern |
Bare except, eval, hardcoded secrets | 5 |
Rule Lifecycle
Gap detected → Group similar gaps → ≥2 occurrences?
│
Yes ▼
Generate Rule (via AI)
│
▼
Deploy to HF Dataset
│
▼
Inject in future prompts
│
▼
Track effectiveness
│
Score < 15%? → Deactivate
How to Use These Rules with Any AI
Step 1 — Generate rules from your conversations
- Upload your Claude Code session files in the 📥 Import Sessions tab
- Click 🔁 Force Re-analyze All in the 🔍 Analysis tab to scan all conversations
- The system detects gaps and calls
Qwen/Qwen2.5-72B-Instructto generate guardrail rules
Step 2 — Export the system prompt
Go to 📋 Rules → click Generate System Prompt → copy the output.
Step 3 — Apply to any AI
Paste the system prompt into:
- Claude — Project instructions or system prompt in Claude.ai
- ChatGPT / OpenAI API —
systemmessage in the messages array - Any API —
{"role": "system", "content": "<paste here>"} - Claude Code — Add to
CLAUDE.mdin your project root
Auto-export from Claude Code sessions
The Stop hook auto-exports sessions when a session ends. Set HF_TOKEN in your shell:
export HF_TOKEN=your_hf_token
# Now every Claude Code session auto-uploads to the dataset on exit
Fetch rules programmatically
from huggingface_hub import hf_hub_download
import json
path = hf_hub_download("vooom/AI_Rule_Learning", "rules.jsonl", repo_type="dataset", token="your_token")
rules = [json.loads(l) for l in open(path) if l.strip()]
active = [r for r in rules if r.get("is_active")]