How to Install Grok Bot Templates: 3 Shared Templates Tested Live

Grok bot templates went live about four or five days before I recorded this. Since then I have been watching people share their bots across X, and three of them caught my eye. So I installed all three, live, on my actual Grok bot setup, and tested them against real work.

Here is what happened with each one. The short version: one is on probation, two earned a spot on the team, and none of them are reinventing the wheel.

What a Grok bot template actually is

A template is someone else's pre-built bot, packaged so you can install it in one click. It carries the bot's instructions, and it can also carry its own skills, memories, integrations, and model rules.

I did not realize that last part until I installed the first one. Creators can bundle separate skills per template before they share it. That changes how much a template can do out of the box.

How to install a Grok bot template

This is the easy part. In the first case, the post had no link at all. I just clicked the template card.

Grok bot detected that I had it installed locally and offered to open with Grok bot. From there the template triggered its own setup prompt and started configuring itself. No copying instructions, no pasting slugs by hand.

If you have Grok bot installed, the share card does the work.

Template 1: The video editor

The template description reads: "Cuts, recuts, shot notes, and edit help from official and owner footage. Diagnoses the treatment first, then finishes a review-ready master. Never auto posts or overwrites sources."

That last line is the kind of guardrail I look for. And it was not just marketing. The bot's own instructions treat camera originals, audio masters, project files, and graphics as read-only, and it works one bot with no delegation.

### The four bundled skills

While the bot was working, I read through every skill it shipped with:

• Video Post-Production Director. Use when turning raw camera, screen, audio, and graphics into a finished video with an edit.
• Viral Short-Form Producer. Turns footage the owner already has into short videos. Its rules include "do not promise virality," originals are read-only, work from proxies and versioned drafts, and never fabricate a quote.
• Per-Clip Scene Research. Writes one sentence on what the clip is about, who wants what, and what the audience must know. Then it opens a browser and searches film grab and stock sources. It does not try to generate footage. I was genuinely shocked by that restraint.
• Film Director Craft. "You are the director. You do not generate or cut from adjectives. A shot is a one camera event." This one is very detailed, with craft numbers and when-to-use rules.

### The 200 MB wall

I tried dragging our most recent live stream into the chat. The file was 1.85 GB. Nothing happened at first, then an error came back: max 200 MB.

I had a feeling this would happen. Dragging a couple gigs of video into chat does not make sense, and it would burn through usage anyway. Your options right now are to put the file in Google Drive and have the bot go access it (untested), or drag in a clip under 200 MB.

### What it found without the file

Instead of uploading, I gave it a simpler instruction: go to Clearmud's YouTube channel, check the most recent live stream, and work your magic.

It came back with four beats: the cold open, the simplicity thesis, the viral payoff, and the "boy who cried wolf" beat.

The cold open was real. On our last live stream I started the broadcast but never actually hit go live. I talked to myself for three minutes before I noticed. The bot found that on its own, which told me it was reading the actual content and not guessing.

### Room for improvement

The rendered clip was incredibly zoomed in. I was almost cut out of the frame. Playback was also laggy and kept buffering. I am honestly not sure yet whether the fix belongs in the skills, the instructions, or the overall workflow.

Listen, it is not perfect, but it works. Whether we keep this one is still an open question.

Template 2: Dr. Eggbot

This one comes from potatoes, aka Lauren, who is on the team over at SpaceX. The description: designs high-quality Grok bots, asks a few preference questions, then creates them with create agent. Every bot gets the same tightness, coding bot or not.

The install showed me something new. This template shipped with memories already loaded into context, an integrations list, and a P stack model rule. "Inherit parent" means each P stack rule runs on the chat's model. I accepted the default to make my life easier.

### It audited my entire fleet immediately

Once it was in, it offered to check up on my bots and routines. I said yes.

The report came back fast:

• 31 bots
• 9 routines
• 16 flags
• Nothing denser than hourly

Then the specifics: a standing wake that re-reads a fat transcript when live, and noisy empty runs it wanted to move off.

I skipped every suggested change. I have things paused right now on purpose because I am mid-test, and I do not want an outside bot editing routines while that is happening. But the audit itself is super nifty. You can review your whole fleet and optimize it from one place. Thank you, potato.

### Does it replace my Staffing bot?

I asked it directly. I already have a Staffing bot, and I wanted a clean split of responsibilities instead of two bots quietly overlapping.

Its answer was blunt. Staffing's job is bigger and sloppier: interview you, draft name and instruction and schedule and connections, create bots, tighten the noisy ones, keep a living roster. Elon is the front door and routes staffing asks there. Hard stops include no delete without my say-so, no X, no mail.

Then it graded Staffing by its own rubric: padded. The base instruction is copy-pasted onto every bot. The "current focus" line still says Pulse is the next new bot, which was already built.

I did not know it was that messy.

On the replacement question: "Only to create. That's my responsibility. I don't own the roster, Elon routing, or tightening the noisy ones. If you talk to both of us about a new bot, you'll get two creators." Its fix was to pick one owner for create so we do not double hire.

I am going to trust potato. Dr. Eggbot stays. The team is now Staffing for the org chart and schedules, Dr. Eggbot for creating bots, and Elon as the front door.

Template 3: Shepherd

This is the one I was most excited about. Shepherd was built by the Herder team. Herder (herdr) is an open-source runtime for coding agents. Shepherd orchestrates agents inside it, and if Herder is not installed, it installs it.

The reason this matters: you stop relying on your chat bots to do heavy codebase work. You spin up Herder and let it handle that.

### Getting it running on the right machine

Install was one prompt. It asked to install the Herder CLI from herder.dev on Grok bot's computer, and I set that to always allow. It then started the server and began wiring Grok, Claude, and Codex, because we already have an engineering team built out.

There was one misunderstanding worth sharing. My engineering bots were set up through Peekaboo on my macOS, so Shepherd initially assumed it should run there. I corrected it: I want Herder on the remote PC, not my Mac. If my Mac goes offline, Herder goes offline. The whole point is an always-on machine.

Once that was clear it installed all the CLIs on the remote computer. Claude Code went in, Claude and Grok were in, then it finished Codex. It kicked off every login so I could authenticate each one.

### Watching it work

I asked how to monitor progress. The answer: click Shepherd in the chat header, hit Command Shift I, click the live computer, and go full screen. That gave me a visual window into exactly what Claude and Codex were doing.

For a test prompt I asked it to research the top ten Grok bot templates being shared on X, from the perspective of Codex, Claude, and Grok, and come up with at least one template idea from each model. It spun up Grok in the CLI right away, with Claude Code close behind.

We could have done all of this manually. This makes it much easier. Shepherd moved up to the engineering team.

Where things landed

| Template | Creator | Verdict | |---|---|---| | Video editor | Shared on X | Undecided. Real beats, but zoomed-in output and laggy playback. | | Dr. Eggbot | potatoes (Lauren) | Kept. Fleet audit plus a clean ownership split. | | Shepherd | Herder team | Kept. Runs Herder on the always-on remote PC. |

There is one more possibility I am watching. If Shepherd covers what my three engineering bots do, I might condense three bots into one and save a lot of usage. That needs a proper side-by-side: one task to Shepherd, one to our Codex agent, one to a Grok-powered bot. That is its own 30-to-60-minute video.

What to take from this

• Read every bundled skill before you trust the bot. Look for read-only rules on sources, no auto-post, no fabricated quotes, no promised virality.
• Stay under 200 MB per upload. Cut a proxy clip or point the bot at cloud storage. Do not expect multi-gigabyte drag-and-drop to work.
• Give the editor a channel, not a file. A YouTube channel and a stream name were enough to find real beats.
• Use a bot-designer template as an auditor. A count of bots, routines, and flags plus a list of noisy runs is worth the install on its own.
• Ask new bots what they own. Assign one owner per responsibility so you do not double hire.
• Put orchestration on the always-on machine. Anything that dies when your laptop sleeps defeats the purpose.

I am not an AI expert. I am building in public and sharing what actually works. If you want a video on a specific topic, tell me who you are, what you do, who your audience is, and what your question is.

Watch the full test here: https://www.youtube.com/watch?v=ScjMoBdzv5c