How the Grok Bot Team Uses Bots for Design, Engineering, and Everyday Work

Grok Bot team members Peng Zheng and Lauren Tan show Peter Yang the bots they use for personal tasks, design and engineering. Their examples range from a chief of staff and a design assistant to a bot that audits other bots. Granola sponsors the episode; Riverside is also listed as a sponsor.

Peng organizes bots around work and life and uses a chief of staff when unsure where a task belongs. One example combines ordering 3D-printing filament with updating a Notion inventory. Another combines researching a used microphone’s price, drafting a marketplace listing and scheduling price reductions.

For email and calendar work, Peng asks bots to identify action items or manage events. When an account needs authentication, the bot can request that he sign in. His writer bot polishes messages, with a draft available for him to review and send; his designer bot helps generate and discuss alternatives.

For a larger idea, Peng puts product, design and engineering bots in a group chat. He demonstrates a hypothetical cat-sound app: each bot contributes its perspective, and a specific bot can be addressed directly. The named bots retain reusable roles and memory while separate chats organize individual projects.

A podcast bot turns articles into audio for listening. In a personal website project, Peng sends a photo and information about a place; a workflow creates miniature-style imagery, removes backgrounds and updates the site. A private-repository playbook guides the steps, making the bot an interface to a publishing pipeline.

Peng still works across design tools and prototypes. His designer bot can use a Figma connector and a skill describing the design system. Sometimes he creates the system and one key frame himself, then asks the bot to extend that direction across the rest of a flow.

Lauren shows engineering workflows, including using an agent to interact with a Linux virtual machine on her computer while working on product support. She emphasizes giving agents a way to verify their work. Access to the local machine depends on the permissions granted.

For her plugin and skill stack, Lauren describes an evaluation playbook that runs tests with different agents or models before a change is accepted. Her goal is evidence that a skill change achieves the intended improvement, rather than relying on a persuasive response alone.

Her Dr. Eggbot creates and audits bots. It can inspect conversations and routines, propose role or skill changes, and flag schedules that wake agents too frequently. Lauren notes that frequent routines can consume substantial usage.

Lauren’s personal assistant Tsuki helped plan a complicated work trip, comparing flights and hotels. She says she asked it to check with her before booking and approved the proposals. She also describes reminders and routine personal purchases; these are examples of her own configured workflow.

For engineering, a chief-of-staff bot talks to an engineering lead, which breaks down work and delegates to other bots. Those bots can start cloud agents and choose different models or reasoning levels for different tasks. Lauren uses this hierarchy to coordinate larger projects.

Lauren says that, in her particular agent-friendly codebase and workflow, some pull requests can merge before she personally reads them. She calls her quality-at-scale aspiration a “Michelin kitchen.” This is a description of her setup, not evidence that unattended merging is universally suitable.

The team also packages working practices into reusable skills, including individual “modes.” Lauren says plugins can combine skills with other tools and that some can be used across Grok Bot and Cursor. The goal is to let more people contribute effectively to the codebase.

A social bot checks mentions and groups recurring bug reports. Lauren can ask an engineering bot to consult it and propose a fix, passing context between roles rather than copying every detail herself. The value, she says, comes from connecting those separate workflows.

Peng describes the product idea as moving from manually doing tasks to delegating them and eventually designing systems and guardrails. Named agents with their own roles are intended to make that interaction more intuitive; the conversation discusses collaboration as a direction, not a detailed product commitment.

Asked how to trust a bot with longer tasks, Lauren recommends observing it, correcting mistakes and turning the improved process into a reusable skill. Once a workflow works consistently, she may automate it as a routine. Peng similarly suggests starting small, testing boundaries, monitoring usage and gradually connecting more tools.