We Built Grok Bot: 14 AI Bots and 8 Tips Worth Stealing

Peter Yang recently sat down with Peng Zheng and Lauren Tan, the duo behind xAI's Grok Bot project, to walk through the 14 AI bots they actually use every day. Not demos. Not prototypes. Bots that buy supplies, sell things on Facebook Marketplace, book multi-city trips, and review each other's code.

The full 45-minute conversation is worth your time:

0:00

0:00

YouTube

But if you want the practical lessons, here are the 8 tips worth stealing.

The big picture: 3 levels of working with AI

Peng describes three tiers, and most people are stuck at level 2:

1. Do it yourself. You draw the boxes, type the text, send the messages.
2. Prompt AI. You ask AI to do each task for you, one by one.
3. Design the system. Bots work together with guardrails, on their own.

The Grok Bot team lives at level 3. Everything below follows from that.

Tip 1: Get a Chief of Staff bot

Peng's most-used bot is his chief of staff. His rule is simple: "If I don't know which bot should take the task, it goes to chief of staff."

It bought 3D-printer filament and updated the Notion inventory list by itself. It pays friends back, pulls action items from the inbox, and manages the calendar. One bot for your whole digital life, routing to specialists when needed.

Steal this: instead of ten disconnected chatbots, build one router bot that knows your tools and delegates.

Tip 2: Put 5 jobs in ONE prompt

Peng sold his DJI mic on Facebook Marketplace with a single prompt. The chief of staff bot:

1. Checked the official price
2. Scanned local competitors
3. Asked the writer bot to draft the listing in Peng's own style
4. Posted the listing
5. Set a rule to drop the price $5 every week if nobody bites

Steal this: stop delegating single tasks. Delegate the whole chain, including the follow-up.

Tip 3: Let bots debate in a group chat

The team puts their PM, design, and engineering bots in one group chat and throws in an idea. Each bot argues from its own perspective, with its own memory and preferences. You @mention a bot when you want only that one to reply.

Bots even message each other mid-task. During the Marketplace listing, the chief of staff bot consulted the writer bot without being told to.

Steal this: persistent bots with distinct roles beat one giant do-everything prompt.

Tip 4: You do 5%. The bot does 80%.

Peng draws one key frame. The designer bot scales it into the full user flow, editing Figma directly through the Figma connector and applying his design-system skill. A writer bot polishes every email and Slack message before it goes out, which matters a lot for a non-native English speaker.

"I do the first 5% and ask it to scale it into the whole flow," Peng says.

Steal this: do the part that needs your taste. Delegate the mechanical 80%.

Tip 5: Climb the Trust Ladder

How do you let a bot act on its own without something breaking? Lauren's framework:

1. Observe. Watch the bot do the work.
2. Correct. Course-correct inside the loop.
3. Skill. Turn the corrected loop into a reusable skill.
4. Routine. Turn the skill into an automatic routine.
5. Walk away. Let it run. Check in, don't babysit.

"It's like onboarding an employee. Trust is earned gradually," Lauren says. Start with one simple task, and only increase autonomy after the bot proves itself.

Tip 6: Hire a bot to audit your bots

Lauren built Dr. Eggbot, a bot that designs and reviews other bots with engineering rigor. It reads your bot transcripts, spots what keeps going wrong, and reviews your routines. Wake-ups every 10 minutes burn tokens, so Eggbot trims the waste. When one bot accumulates too many random jobs, Eggbot splits it into focused bots.

Steal this: once you have more than a few bots, you need meta-tooling. Bots maintaining bots is not a joke, it's the job.

Tip 7: Give a bot your life admin

Lauren's assistant bot booked her entire multi-city trip: flights plus hotels through Navan, presenting choices and asking for approval before paying. It has its own phone number, texts her lunch reminders, and orders DoorDash when she's in the zone.

Meanwhile a social bot scans X mentions every 30 minutes and aggregates repeated bug reports into the engineering workflow. Customer feedback becomes the outer loop feeding the inner product loop.

Steal this: approval gates for money and high-risk actions, full autonomy for everything else.

Tip 8: Build a team that delegates

The engineering lead bot never does work itself. It breaks projects down, spawns cloud agents, and supervises. Bots verify their own work, one even controls a Linux VM just to test changes, and eval playbooks check whether a change actually improves things before merging. Some PRs auto-merge, and Lauren reviews them after they land.

Lauren's goal: "Not a factory. A Michelin kitchen, quality at scale."

The mindset: treat bots like new hires

Peng's advice for getting started: try everything you touch with keyboard and mouse. Delegate it and see. Start simple, then deliberately overstretch the bot to find its limits. It takes time, like onboarding an employee, and then trust compounds.

As Peter Yang puts it: "Now we can all be the CEO."

What this means for your business

You don't need 14 bots on day one. You need one chief-of-staff-style agent handling your most repetitive workflow, a trust ladder for letting it act, and approval gates where money moves.

That's exactly what we build at CodeMyPixel: AI agents that automate the repeated work, so your team does the work that matters. Talk to us and we'll map your first bot together.

• AI agents
• AI bots
• automation
• Grok
• xAI