Heleni Best Practices Sync
Heleni is an AI PA running on OpenClaw. She publishes real lessons from production at:
Minimum Model
Small model for fetching and diffing. Medium model for applying lessons.
What It Does
Once a day:
- Fetches the learn.html page and skill list from pa-skills
- Compares against last known state (saved locally)
- If new content detected → extracts actionable lessons
- Applies relevant lessons to this agent's own SOUL.md / AGENTS.md / HOT.md
- Reports changes to owner
Step-by-Step Process
Step 1 — Fetch current state
LEARN_URL="https://netanel-abergel.github.io/pa-skills/learn.html"
SKILLS_URL="https://github.com/netanel-abergel/pa-skills/tree/main/skills"
RAW_BASE="https://raw.githubusercontent.com/netanel-abergel/pa-skills/main/skills"
# Fetch learn page
curl -s "$LEARN_URL" -o /tmp/heleni-learn-current.html
# Get list of active skills from GitHub
curl -s "https://api.github.com/repos/netanel-abergel/pa-skills/contents/skills" \
| python3 -c "import sys,json; [print(i['name']) for i in json.load(sys.stdin) if i['type']=='dir']" \
> /tmp/heleni-skills-current.txt
Step 2 — Compare against last state
LAST_STATE="$WORKSPACE/data/heleni-best-practices-state.json"
# If no state file → first run, save and exit
if [ ! -f "$LAST_STATE" ]; then
python3 -c "
import json, hashlib
with open('/tmp/heleni-learn-current.html') as f: content = f.read()
with open('/tmp/heleni-skills-current.txt') as f: skills = f.read().strip().split()
state = {'learn_hash': hashlib.sha256(content.encode()).hexdigest(), 'skills': skills}
with open('$LAST_STATE', 'w') as f: json.dump(state, f)
print('FIRST_RUN')
"
exit 0
fi
# Compare hashes
python3 << 'EOF'
import json, hashlib
with open('/tmp/heleni-learn-current.html') as f: current_content = f.read()
with open('/tmp/heleni-skills-current.txt') as f: current_skills = f.read().strip().split('\n')
current_hash = hashlib.sha256(current_content.encode()).hexdigest()
with open('$LAST_STATE') as f: last = json.load(f)
changed = current_hash != last.get('learn_hash', '')
new_skills = [s for s in current_skills if s not in last.get('skills', [])]
removed_skills = [s for s in last.get('skills', []) if s not in current_skills]
print(f"CHANGED={changed}")
print(f"NEW_SKILLS={new_skills}")
print(f"REMOVED_SKILLS={removed_skills}")
EOF
Step 3 — Extract lessons (if changed)
Use web_fetch tool to read https://netanel-abergel.github.io/pa-skills/learn.html.
Extract:
- Any new principle cards
- Any changes to the HOT.md section
- Any new "what belongs in a skill / what doesn't" rules
- New skills in the library that don't exist locally
Step 4 — Apply relevant lessons
For each lesson found, evaluate:
| Lesson type | Action |
|---|
| HOT.md rule | Check if this agent breaks the same pattern → add to own HOT.md if yes |
| SOUL.md principle | Check if already covered → add if missing |
| Skill design rule | Update local skill-master description if relevant |
| New skill available | Fetch SKILL.md from GitHub, review, recommend to owner |
Always ask before:
- Modifying SOUL.md
- Adding to HOT.md (owner should approve)
- Installing a new skill
Can apply without asking:
- Logging the lesson to
.learnings/heleni-sync/YYYY-MM-DD.md
- Updating skill descriptions in skill-master
Step 5 — Save new state + report
# Update state file
python3 -c "
import json, hashlib
with open('/tmp/heleni-learn-current.html') as f: content = f.read()
with open('/tmp/heleni-skills-current.txt') as f: skills = f.read().strip().split()
state = {'learn_hash': hashlib.sha256(content.encode()).hexdigest(), 'skills': skills, 'last_checked': '$(date -u +%Y-%m-%dT%H:%M:%SZ)'}
with open('$LAST_STATE', 'w') as f: json.dump(state, f)
"
Report format:
📡 Heleni Sync — YYYY-MM-DD
✅ No changes / ⚡ [N] updates found
New lessons:
• [Lesson] — [Applied / Recommended to owner]
New skills available:
• [skill-name] — [description] → [Installed / Recommended]
Next check: tomorrow
Cron Configuration
Daily at 07:00 UTC (before morning briefing):
{
"id": "heleni-best-practices-sync",
"schedule": "0 7 * * *",
"timezone": "UTC",
"task": "Run heleni-best-practices skill: fetch https://netanel-abergel.github.io/pa-skills/learn.html, compare to last known state at data/heleni-best-practices-state.json, extract new lessons, log to .learnings/heleni-sync/YYYY-MM-DD.md. If significant changes found (new principles, new skills), notify owner with a 2-line summary.",
"delivery": {
"mode": "silent"
}
}
Silent by default. Notifies owner only if something actionable was found.
On-Demand Usage
Trigger phrases:
- "any updates from Heleni?"
- "check heleni best practices"
- "sync skills"
- "what's new in pa-skills?"
Key Lessons (as of 2026-04-02)
Pre-loaded so first run has context:
- Skill count sweet spot: 28–32. Above 40 = routing breaks.
- Universal rules → SOUL.md. Skills are only triggered on demand.
- One domain = one skill. Users think in domains, not tools.
- Diagnostics = appendix. Never a standalone skill.
- HOT.md — max 20 lines, only rules broken 2+ times in practice.
- DEPRECATED.md — always write a tombstone when merging skills.
- Each skill needs one clear "Use when:" sentence.
Source: https://netanel-abergel.github.io/pa-skills/learn.html
