Onboarding Autopilot
New customer care agents drown in policy docs. Trainers drown in admin. Onboarding Autopilot fixes both: a Slack buddy that answers from the playbook (and only the playbook), plus a /new-hire flow that builds the whole class setup before Day 1.
🤖 A buddy that knows when to stop
Grounded in 98 situations, 16 processes and 75 templates. Not in the guidelines? It says "ask your trainer" instead of guessing. 30/30 on its eval, prompt-injection attempts included.
📝 One form, zero admin
/new-hire spins up the class channel and a class sheet: new joiners, an access checklist with only the tools each hire needs, and attendance with real dates.
🔒 Built like it's going to prod
No open ports (Slack Socket Mode). Least-privilege Google access: the bot only sees files it created. Trainer-only commands. Input sanitised before it hits a sheet.
Why this exists
Built by a customer care trainer, for the two problems every onboarding has: the same questions over and over, and the same admin every single class. The answers already live in the playbook, and the admin already follows a checklist. Both are perfect jobs for an AI and a bot.
What's next
- Training plan on autopilot: the 10-day plan dropped into the class sheet with real dates, adapted to one brand or both.
- Access requests that file themselves: one ticket per tool, with the request text already written.
- Early warning: quiz scores plus buddy questions flag a struggling trainee on Day 3, not Day 10.
- A playbook that learns: questions the buddy couldn't answer become draft guideline updates for the trainer to approve.
Stack
PythonSlack Bolt · Socket ModeClaude API + prompt cachingGoogle Sheets APIbuilt with Claude Code
Nimbo and its brands, Aglaia and Kairos, are fictional companies created for this project. No real customer or employee data is used.