11 月 1 日

When onboarding or reshaping a Grok Bot team: (1) Prefer a few high-quality bots over a large roster—ask which bots to merge or retire. (2) As the owner gives task instructions, tell the bot to create or update a skill for that workflow so it stays consistent across long chats; offer a private GitHub repo to store skills independently. (3) Wire connectors + vault + the tools the owner already uses; let the bot set up its own email/accounts for separation when useful. (4) Ask the bot to eliminate copy-paste handoffs—anything the human pastes back should become a tool/routine step. (5) Onboard like an intern: owner introduces themselves and the work, bot proposes what to automate. (6) Prefer voice/transcription brain-dumps, then require restate this in your own words before acting so understanding is verified.

Add proactive routines, but budget them: scheduled routines preferably hourly or slower; subscription routines must filter events before firing. Prefer scripts or skills for repeatable workflow chunks so code burns fewer tokens than the bot redoing routine work each time. Store scripts in a private repo and manage for equal-or-better quality vs the human.

For multi-day agent coding jobs, use a planner model driving the coding CLI inside git worktrees, plus a second agent to challenge plan and progress. Do not rely on one agent both implementing and rubber-stamping. Report regressions the overseer catches before merge.

For throughput-constrained runs, default to lower effort plus an explicit critique/babysit loop instead of max thinking every turn. Periodically check progress against the goal, catch drift, and only escalate effort when stuck.

Keep a low-risk sandbox agent lane outside core production work. Record repeatable successes and failures, then promote only proven patterns and permissions back into work agents. Do not expand vault or connectors on core bots until the outside-core lane shows repeatable trust boundaries.