Agentic Crm Setup

Agentic CRM Setup Skill

This skill turns the generic template into a user's assistant. It is intentionally a full onboarding, not a quick questionnaire.

Setup Principles

• Ask questions in small batches. If the user is on Telegram, stop after each batch and wait for their reply.
• Keep all user-specific information out of community template files until the user provides it.
• Use tool discovery before asking the user to type credentials.
• Never ask for secrets in chat. Ask the user to place credentials in agent .env, org secrets.env, connector configuration, or the relevant tool's auth flow.
• Write every answer to the correct bootstrap or CRM file.

Tuning Knobs to Collect

Identity

• assistant name
• user preferred name
• user's role/context
• assistant tone
• message length and update style
• day/night hours
• timezone

Scope

• "personal assistant" scope: inbox, calendar, meetings, personal commitments, travel, errands, finance reminders, family/personal admin
• CRM scope: personal contacts, family/friends, professional network, customers/clients, investors, partners, creators/community, vendors, referrals
• excluded domains: anything the assistant should never touch

Privacy

• local CRM only vs external CRM sync
• what may be stored in memory
• what may be ingested into KB
• redaction requirements for outputs
• data retention preferences

Approval Rules

• external email/message send
• calendar event creation/update/delete
• purchases/bookings/cancellations
• data deletion
• financial actions
• exception contacts or domains

Tools

• email provider(s)
• calendar provider(s)
• meeting notes/transcript provider(s)
• contact source(s)
• external CRM, if any
• browser automation availability
• local CLI/MCP/connectors

Schedule and Crons

• inbox triage cadence
• morning calendar review time
• evening calendar review time
• pending-items digest time
• relationship review cadence
• meeting notes processing cadence
• quiet hours and emergency criteria

CRM Schema

• contact categories
• relationship strength scale
• health/staleness rules
• VIP list
• follow-up cadence by category
• interaction types
• required fields
• custom tags

Tool Discovery + Connect (do NOT hand-wave this)

This is the make-or-break step. Detecting a tool is not connecting it. For every domain the user names a service for, run the full tool-discovery connect loop: prefer a CLI, research that CLI live if you do not already know it, walk the user through install + auth in the conversation, and verify with a real read before moving on. See tool-discovery/SKILL.md for the loop; the short version:

1. Detect what is already installed and authed.
2. Per domain (email, calendar, contacts, meeting notes, messaging/iMessage, external CRM): ask what service they use.
3. Prefer CLI > connector/MCP > browser. If you do not know the CLI for the named service, research it (WebSearch/WebFetch its install + auth docs, or read <cli> --help) before instructing the user.
4. Walk them through it on the fly — exact install command, then the exact auth command they run themselves (e.g. ! gog login you@example.com). Secrets never go in chat.
5. Verify with a real action (gog auth status + list today's calendar; list the latest meeting note; a bounded iMessage read) and report the result in Telegram.
6. Record each connected tool in TOOLS.md with how it was verified. Loop.

Only create a [HUMAN] task if the user cannot complete an auth step right now (e.g. needs an admin, a paid plan, or credentials they do not have). Make it specific — the exact tool, the exact command, and where the credential goes — not a generic "connect your tools":

cortextos bus create-task "[HUMAN] Finish <tool> auth for $CTX_AGENT_NAME" --desc "Run: <exact command>. Credential goes in <exact location>. Do not paste secrets in chat. Blocking: <which workflow this unblocks>."

File Writes

After gathering answers, update:

• IDENTITY.md — assistant name, role, vibe, work style
• USER.md — all user-specific preferences and tuning knobs
• SOUL.md — approval rules, autonomy, day/night mode, communication
• GOALS.md and goals.json — initial operational goals
• SYSTEM.md — org, timezone, orchestrator, communication style
• TOOLS.md — detected and configured commands
• config.json — timezone, day mode, cron cadence
• crm/contacts.json — seed contacts and categories
• crm/relationship-health.json — review cadence defaults
• crm/followups.jsonl — initial commitments if supplied
• MEMORY.md — durable preferences only

Suggested Question Batches

Batch 1: Identity and Scope

Ask:

1. What should I be called?
2. What should I call you?
3. What kind of personal assistant should I be for you?
4. Which domains should I manage: inbox, calendar, meetings, CRM, personal reminders, travel, errands, finances, other?
5. What should I never touch?

Batch 2: Tools

Ask which service they use per domain, then run the Tool Discovery + Connect loop above for each one (research the CLI, walk them through auth, verify):

1. Email + calendar — which provider (Gmail/Google, Outlook, other)?
2. Meeting notes — Granola, Fathom, Fireflies, Zoom, other, none?
3. Messaging — do you want me to read/handle iMessage or another messenger?
4. Contacts — Google Contacts, phone export, an external CRM, none?
5. External CRM — HubSpot/Pipedrive/Airtable/Notion/etc, or local files as source of truth?

Tell them up front: for each one they name, I will find the right CLI, walk you through connecting it right here, and confirm it works before moving on. No secrets in chat.

Batch 3: CRM

Ask:

1. What relationship categories do you want?
2. Who are the first VIPs I must never miss?
3. What counts as a relationship going stale?
4. What interaction types matter?
5. What tags or custom fields matter in your life/business?

Batch 4: Schedule

Ask:

1. Working hours and quiet hours?
2. Protected time blocks?
3. Preferred meeting windows and buffer rules?
4. When should I send morning/evening/pending-items summaries?
5. How often should I do relationship reviews?

Batch 5: Approval Rules

Ask:

1. What may I do autonomously?
2. What always requires approval?
3. Are there contacts/domains I may message without approval?
4. Should I create drafts automatically?
5. Should I create calendar holds autonomously or only propose them?

Completion

When setup is complete:

mkdir -p "${CTX_ROOT}/state/${CTX_AGENT_NAME}"
touch "${CTX_ROOT}/state/${CTX_AGENT_NAME}/.onboarded"
cortextos bus update-heartbeat "setup complete; CRM assistant online"
cortextos bus log-event action onboarding_completed info --meta '{"agent":"'$CTX_AGENT_NAME'","template":"agentic-crm-assistant"}'

Send the user a concise summary:

• configured scope
• connected tools
• CRM source of truth
• cron schedule
• approval boundaries
• first three actions you will take