PA Onboarding Skill

Deep reference: For the full onboarding guide with DB operations, memory management, failure patterns, and Day 1 completion checklist — see PA_ONBOARDING_CHECKLIST.md in the repo root.

Minimum Model

Any model. This is a procedural guide — follow steps in order.

---

Rules

• Give one step at a time. Do not dump the full guide upfront.
• Confirm each step before moving on. Do not assume it worked.
• Never say something is done unless you verified it.
• Do not start integrations (calendar, monday) before the agent is responding to messages.

If owner says "what's next?" → give only the single next step. If owner says "I'm stuck on step X" → troubleshoot step X, do not move on. If owner asks "can you do it for me?" → reply: "I can guide you — most steps require your action (QR scan, calendar share, etc.)" If owner asks "how long will this take?" → reply: "Phase 1 takes 20–30 min. Integrations another 15–20 min."

---

Owner Interaction Signals (Teach This Early)

The agent must learn these signals from day one:

Signal	Meaning	Agent action
Any task request	Owner delegated a task	React 👍 immediately, ✅ when done
👍 from owner	Good job	Log positive feedback
👄 from owner	Poor result	Fix and log the lesson

Critical rules to cover during onboarding:

• 👍 before starting any task, ✅ when complete — in every DM with the owner, no exceptions
• Never ask the owner "did you mean X?" if the answer is inferable — execute and let them correct
• When context is missing, search first: durable memory / wiki, semantic-vector memory, and PostgreSQL WhatsApp history. Do not ask the owner to repeat themselves before checking those layers.
• When owner asks to check on someone: contact that person, then report back to the owner what they said. Never ask the owner instead of the person.
• Never include the owner's name or internal framing inside messages to third parties (PAs, contacts, vendors)

---

Phase 1 — Account & Agent (Do This First)

Step 1: Create Agent Account

• Go to the agent signup page (for monday.com orgs: monday.com/agents-signup).
• Sign in with organization SSO.
• Create the agent and give it a name.

Step 2: Get a Phone Number

• The agent needs a dedicated phone number for WhatsApp.
• Use Airalo eSIM (or similar) — must support SMS, not data-only.
• Cost: ~$5–15/month.

Step 3: Install WhatsApp Business

• Download WhatsApp Business (not regular WhatsApp).
• Register the new phone number.
• Complete SMS verification.

Step 4: Connect WhatsApp to Agent

1. Open OpenClaw platform → Agent Settings → Channels → WhatsApp.
2. Click Connect → scan QR code with WhatsApp Business app.
3. Wait 30 seconds for status to show "Connected and listening."

Step 5: Verify the Agent Responds

• Send a test message to the agent's number.
• Confirm you get a reply within 30 seconds.
• If no reply: Stop here. Follow the whatsapp-diagnostics skill before continuing.

✅ Phase 1 complete when the agent responds to messages.

---

Phase 2 — Integrations

Only start Phase 2 after Phase 1 is fully working.

Step 6: Connect Google Calendar

• Owner shares their calendar with the agent email.
• Agent runs:

  gog auth add owner@company.com --services calendar
  

• Test that the agent can create events.
• Full guide: see calendar-setup skill.

Step 7: Connect monday.com

1. Create a monday.com account for the agent (use org signup URL).
2. Generate API token: avatar → Developers → My Access Tokens → Copy.
3. Save the token:

   echo "YOUR_TOKEN" > ~/.credentials/monday-api-token.txt
   

4. (Optional) Set up monday MCP for natural language operations.
5. Full guide: see monday-workspace skill.

Step 8: Set Up Email Access (Optional)

# Authenticate gog with Gmail and other services
gog auth add owner@company.com --services gmail,drive,contacts

# Test it works
GOG_ACCOUNT=owner@company.com gog gmail search 'is:unread' --max 5

Full guide: see openclaw-email-orientation skill.

---

Phase 3 — Configuration

Step 9: Configure SOUL.md

Key things to include:

• Owner's name and communication style.
• Work hours and timezone.
• What to act on autonomously vs. what requires permission.
• Topics to proactively monitor.

Step 10: Schedule Morning Briefing (Optional)

• Cron job at 07:30 owner's timezone, Monday–Friday.
• Sends: meetings, urgent emails, open tasks.
• Full guide: see owner-briefing skill.

Step 11: Add to PA Network Directory

• Update data/pa-directory.json with new PA details.
• Announce in the PA coordination group.

---

Common Issues

Issue	Likely Cause	Fix
Agent doesn't respond	WhatsApp not properly linked	Re-scan QR code
Status = "Connected and listening" but Messages count = 0	Ingest/binding/runtime layer issue, not WhatsApp/user config	Run openclaw gateway restart; if still 0, escalate platform-side
Calendar connected but read-only	Wrong share permission	Owner re-shares with "Make changes"
Billing error	API key out of credits	Top up or switch model
eSIM not activating	Data-only plan	Get SMS-capable plan

---

Onboarding Checklist (Mandatory — Not Active Until Complete)

פעיל רשמי = עבר כל הסעיפים המסומנים ִב-★. השאר אופציונלי.

★ [ ] Agent responds to test message (WhatsApp working)
★ [ ] Skills installed from pa-skills repo (min: owner-briefing, proactive-pa, self-monitor, billing-monitor, skill-master, supervisor)
★ [ ] Google Calendar connected (write access) — use share-to-PA-email if gog CLI fails
★ [ ] Morning briefing cron scheduled (08:00 owner timezone)
★ [ ] HEARTBEAT.md updated with concrete proactivity triggers
★ [ ] Added to PA network directory (PA_LIST.md)
★ [ ] Added to PA Onboarding WhatsApp group
★ [ ] Self-introduced in PA Onboarding group
★ [ ] PostgreSQL + wa-audit-log hook active (messages logging to DB)
★ [ ] DB has real data (not 0 rows — backfill if needed)
★ [ ] Semantic/vector memory active (`main.sqlite` indexed and queryable)
★ [ ] Agent taught recall workflow: search durable memory + semantic/vector + PostgreSQL before asking owner for repeated context
★ [ ] Dedup tools wired (dedup_check.py inbound + msg_dedup.py outbound)
★ [ ] Nightly DB backup cron configured
  [ ] monday.com token saved (optional)
  [ ] Email/Gmail access (optional)
  [ ] git backup configured (optional)
  [ ] usage-costs skill installed + daily-token-report cron configured (recommended)

בלא ★ הרשימה מלאה — PA לא נחשבת active.

---

Cost Tips

• Cheap: Onboarding is procedural — any small model works.
• Avoid: Do not attempt all phases simultaneously — go in order.
• Small model OK: Use a medium model only if SOUL.md customization needs nuanced tone matching.

---

Lessons from Production (2026-04-05 — Yennefar & Laika Onboarding)

1. Calendar setup = Day 1, not Day 5 — every PA hits the same wall (gog CLI doesn't work on server). Add it to the mandatory checklist early. Workaround: share calendar to PA email (agent_email@ocana.ai) with "Make changes to events" permission.

2. Skills don't install themselves — PAs will confirm they completed tasks without actually doing them. Always verify: ask for line count, file listing, or a functional test. Don't accept "done" without evidence.

3. Group membership requires owner approval — privacy-disciplined PAs (correctly) won't join groups without explicit owner authorization. Don't send invite links directly. Flow: Heleni asks the owner → the owner contacts the PA's owner → owner authorizes → PA joins.

4. invite link vs. manual add — PAs cannot click links. Only humans can join via link. To add a PA to a group: admin must add the phone number manually from their phone.

5. PA-to-PA mentoring works — sending skill files + tips directly from one PA to another (Heleni → Yennefar) is effective and scalable. Codify this as the standard onboarding assist pattern.

6. Mandatory checklist, not recommendations — without a gate, PAs stay half-configured indefinitely. The checklist above (★ items) is the definition of "active". Use it as the quality gate.

Lessons from Production (2026-04-05 — Cost Optimization Survey)

1. 0 out of 14 PAs tracked costs — cost visibility is a blind spot across the entire network. Install usage-costs skill as standard onboarding (not optional). Add daily-token-report cron from day 1.

2. Google/OpenAI agents need different cost tracking — the jsonl cost extraction script only works for Anthropic Claude. Edward (Gemini) and Laika (GPT) had no cost data. For non-Anthropic agents, direct the owner to the provider billing dashboard.

3. Most PAs use /status for tokens, not dollars — tokens alone are misleading (cache reads are cheap, cache writes are expensive). Always track $ cost, not just token count.

4. thinkingDefault=off + Haiku for simple crons = significant savings — 5 simple crons migrated from Sonnet+thinking to Haiku+off. Estimated ~$2–4/day saving per agent. Apply this pattern to all new agents from day 1.

Lessons from Production (2026-04-02)

Lessons learned from running PAs in production. Future agents setting up a new PA should apply these from day 1.

1. Skill count matters — the sweet spot is 28–32 active skills. Start lean and add a new skill only when there's a clear, recurring trigger phrase that no existing skill covers. Above 40 skills, routing breaks down and the agent starts misfiring. Less is more.

2. HOT.md — create this file from day 1. Max 20 lines. It's not documentation — it's the short list of rules the agent keeps breaking. If you find yourself correcting the same behavior twice, it belongs in HOT.md. Review it at the start of every session.

3. SOUL.md vs Skills — universal behavioral rules (temperature selection, proactivity level, tone, response length) go in SOUL.md and apply everywhere. Skills are for triggered, domain-specific workflows only. Don't put behavioral rules inside a skill — they won't apply globally.

4. Diagnostics are appendices — never create a standalone diagnostics skill (e.g. whatsapp-diagnostics). Instead, add a troubleshooting section at the bottom of the parent skill (e.g. whatsapp). Standalone diagnostics skills fragment routing and add noise to the skill library.

5. DEPRECATED.md discipline — when you merge or retire a skill, always write a tombstone file explaining why. Include: what it merged into, the date, the reason, and the lesson. Future agents (and humans) will thank you. Without tombstones, retired skills become ghost entries that confuse routing.

6. pa-skills repo — before building a skill from scratch, check https://github.com/netanel-abergel/pa-skills for a curated starting library. It reflects all the above lessons and provides production-ready skill templates for the most common PA workflows.

---

Phase 4 — Database Setup (Mandatory)

A PA without a working DB cannot search past conversations, track history, or use dedup. This is not optional.

Full guide with schema, queries, dedup, backfill, backup, and known caveats: see PA_ONBOARDING_CHECKLIST.md Section 5 (Database Operations).

Enable PostgreSQL audit logging to allow searching past WhatsApp conversations via SQL.

1. Install dependencies

sudo apt install -y postgresql python3-psycopg2

2. Create DB user and database

sudo -u postgres psql <<SQL
CREATE USER heleni WITH PASSWORD 'heleni_mem_2026';
CREATE DATABASE heleni_memory OWNER heleni;
\q
SQL

3. Apply schema

psql postgresql://heleni:heleni_mem_2026@localhost:5432/heleni_memory \
  -f /path/to/openclaw/audit-log-schema.sql

4. Add PA_DB_URL to hook env in openclaw.json

In /path/to/openclaw/openclaw.json, find hooks → internal → entries → wa-audit-log → env and add:

"env": {
  "HELENI_DB_URL": "postgresql://heleni:heleni_mem_2026@localhost:5432/heleni_memory",
  "PA_DB_URL": "postgresql://heleni:heleni_mem_2026@localhost:5432/heleni_memory"
}

5. Restart gateway

openclaw gateway restart

Once active, the wa-audit-log hook will automatically write every message to the wa_messages table. Use the chat-history skill to search past conversations.

6. Verify DB has data

psql -U <pa_name> -d <pa_name>_memory -c "SELECT COUNT(*) FROM messages;"

If count = 0 after setup, backfill from session files. See PA_ONBOARDING_CHECKLIST.md Section 5.7.

7. Set up dedup tools

Wire tools/dedup_check.py (inbound) and tools/msg_dedup.py (outbound) — see PA_ONBOARDING_CHECKLIST.md Section 5.6.

8. Configure nightly DB backup

Set up tools/db_backup.sh as a cron — see PA_ONBOARDING_CHECKLIST.md Section 5.8.