Grok Bot Bots and context
Each Bot is a durable teammate with its own conversation, description, and developing memory. Vague roles (“General Helper”) waste tokens on re-explaining scope every task.
Hub: Tokenminning in Grok Bot.
Give each Bot one job
Create a separate Bot when the work has a distinct:
- Goal or area of ownership
- Set of tools and sources
- Working style
- Approval boundary
- Recurring schedule
Good jobs: Talent Scout, Expense Manager, Bug Reproduction. Operational descriptions beat category labels.
Example description:
Own the weekly account-health review. Pull product usage and support signals, flag evidence of churn or expansion, and produce a linked watch list. Never contact a customer or change an account without approval.
Description vs conversation
| Field | Put here |
|---|---|
| Description | Standing boundaries, output format, safety rules that persist |
| Message | Task-specific inputs, file attachments, this week’s account list |
Wrong: repeating “never send external email” in every message when it belongs in the description.
Memory is not a database
Bots retain preferences and summaries, but memory is not authoritative:
- Keep changing facts in the source system
- Ask the Bot to cite or reopen current data for consequential decisions
- Correct stale assumptions directly
- Put explicit safety boundaries in the description, not chat alone
Re-fetching live CRM or analytics data costs less than acting on stale memory.
Roster sizing
An account supports up to 50 Bots and group chats combined. Start with the smallest useful roster:
- One Bot owns an end-to-end outcome
- Add another only when a stable specialist role emerges
- Use group chats when the handoff itself must be visible
- Hide inactive Bots instead of deleting if you may need the history
Duplicating a Bot copies profile, skills, and routines — not conversation history or learned memory. Rename and rescope before assigning work.
Bot-to-Bot handoffs
All Bots share one cloud computer. They can message each other and pass context through group chats. Prefer explicit handoffs (“Research complete — files in /workspace/reports”) over re-pasting large summaries into a new thread.