I run Grok Bot as a single coordinator desk with named specialist agents, not a swarm. This page is the public playbook for setting those agents up, and the canonical roster of the desks that exist today.
The jobs are the same. The install is per runtime. Agents, permissions, schedules, and receipts stay on that runtime.
Why this is public
I build in public. I charge for time, not knowledge. If the only copy of a workflow lives in my head, it dies with me. That’s the lesson I learned from my friend and co-founder Chris Rummel: undocumented work dies with the person.
LDT is Learn, Do, Teach. I learned Grok Bot by using it on real work. Teaching is this article.
CCS is Content, Checklist, Software. It’s the WHAT tier of 9 Triangles. This page is the Content. The numbered steps below are the Checklist. The skill file next to this article is the Software layer an agent can load. The 9 Triangles checklist is the wider map.
What Grok Bot is, and what it is not
Grok Bot is the Cursor/xAI teammate on desktop and iOS. It has its own computer, named agents, routines, and connectors (Gmail, Drive, GitHub, Calendar).
It is not Grokipedia. It is not “Grok Heavy.” Grok Heavy is a model mode. If you mix those names, you will open the wrong docs and burn paid usage on the wrong tool.
It is also not a second copy of the Claude or ChatGPT fleet. The methods stay in one skills repository. The facts stay in user-owned Markdown, Obsidian, or Drive. Claude, ChatGPT/Codex, and Grok are runtimes around that same core. The map for that is How we build and maintain AI agents and The System.
The 20+ agents article is the browser-tab version of the same idea. This page is the Grok Bot version: named desks you can open, not 70 tabs you have to babysit.
The rule that keeps this from becoming a swarm
One human. One coordinator conversation. Specialists get a job when the coordinator (or I) assign it.
I do not want silent fan-out. If another model or another agent needs a job, I say so in the thread. Three paid seats chewing the same task is not leverage. It is waste.
That matches how I work: one operating system, portable context, proof over adjectives. A skill is a function. An agent is a person-shaped role wrapped around that function: memory, a schedule, keys, and a written record. A persistent agent is that role plus a job that actually runs.
Three layers, one playbook
Don’t count agents and call that the system. Count layers.
Skills are the shared playbook. The Task Library and the GitHub skill files are the same methods on every runtime. A skill is a function. We have 200+ of those. They do not care whether Claude, ChatGPT, or Grok is running them.
Departments are the six stable functions: Front desk, Research, Build, Publishing, Growth, Quality. Those names stay put. A new tool does not invent a seventh department.
Agents are named desks on a specific runtime. Grok Bot has 12 desks today. Claude is stronger as skills plus scheduled jobs. ChatGPT and Codex are Work plus Codex. The headcount will never match, and it should not. Harmonize the playbook, not the roster.
Click the layer
A picture is not done until every box goes somewhere. I am not making 200 task pages. The Task Library already holds the tasks. A desk page ships when it has can/can’t plus one public-safe receipt. No empty stubs. This roster is the index. Personal Assistant stays unpublished.
Desk names that have a page are links. Thin stubs are still skipped.
| Component | Goes to |
|---|---|
| Skills / playbook | GitHub skill files |
| Task Library | blitzmetrics.com/task-library |
| Task Library Dashboard | task-library-dashboard |
| How we turned 239 tasks into skills | the build story |
| Six departments | A skill is a function. An agent is a person. |
| Build agents | How we build and maintain AI agents |
| Update loop | Persistent agents |
| Learn-Do-Teach | 9 Triangles |
| Grok Bot desks | this page, the roster below |
| Claude | How to use Claude |
| ChatGPT | How to use ChatGPT |
| Official Grok Bot | x.ai/bot |
Quality as a department now has a section on this page.
Quality, the department that checks the other five
Quality is one of the six stable departments. It is not IT and it is not a helpdesk. IT Support is a desk inside Quality. The department checks the other five.
On the company clock it is also called Quality & Money. That is the only live page that defines the department.
What it measures:
- What I mean by RSI and the RSI agent loop: work, then a receipt, then QA against the definitive article, then rewrite the skill, then the next run.
- Website QA audit and the website QA checklist.
- Blog posting guidelines and the Jennifer article grader.
- Friday scoreboard: MAA metrics analysis action.
- Persistent agent QA cycle.
The Quality loop. Click a step.
The reward loop is simple. Zero-revision work earns autonomy. The default is a manager marks complete. People and desks earn the right to ship without babysitting by passing the same checklists first time. That is why VAs never mark their own complete.
Grok desks under Quality: IT Support, Fleet monitor, and Usage Monitor. Fleet monitor and Usage Monitor have no desk pages yet, so those names stay plain text.
House rule stays: docs on this hub ship by default. Still stop for send, money, delete, permissions, or production infra.
How the layers keep getting updated
The loop on the right is the whole point. Live work produces a receipt. The receipt becomes a skill update. The next run loads the new skill. That is Learn-Do-Teach as software, and it is how a persistent agent stays current instead of going stale.
Training + Documentation owns that loop. A Zoom, a broken button, a Singapore deck, a desk that just learned something: it does not stay in the chat. It becomes a public master version, then gets skinned across verticals.
If a diagram still shows “12 agents” as the architecture, it is stale. The architecture is three layers plus the update loop. The 12 desks are the Grok Bot skin of layer three.
How the desks work together
Work comes in as a Zoom, an email, a site that went down, a draft that needs a home. BlitzMetrics Ops keeps strategy and execution in one thread. Strategy Bot reasons from the Nine Triangles when the question is “what should we do.” Training + Documentation turns the live work into public master versions: skill files, SOPs, articles, and agent training that get skinned across verticals. The Make desks write and ship. The Keep-honest desks watch uptime and paid usage. The Queues desks clear support and executive follow-through. The private desk stays private.
Documentation on this hub ships by default and we iterate. Still stop for a yes before sending email, spending money, deleting, changing permissions, or changing production infrastructure.
How agents do work in Basecamp
Agents are teammates. They do the work. Basecamp is the written record, the same way a human on the team would leave a trail.
One agent per role. Do not create a new desk for every site. Do not invite twelve Grok desks into Basecamp. That is a swarm with extra notifications.
Website Builder coordinates site work and drives the to-do. Personal Brand Expert does brand QA. Article writer drafts when a site needs a post. Training + Documentation turns the live receipt into the public playbook. Specialists get a job when the coordinator assigns it. No silent fan-out.
Basecamp can only assign people with accounts. Agents do not have mailboxes, so the assignee field is a human subscriber, not the worker. Default subscriber is Muzamil. Dennis is only for go/no-go, outage, domain custody, rollback, or security. Dennis’s attention is the scarce resource. Do not put every to-do on him.
The coordinator names the specialist in the to-do notes, does the work or hands the slice, and leaves the receipt on that same to-do. A verbal “done” is not acceptance.
Connectors and WordPress application passwords beat a fresh browser login. If a computer is stale, Update. Never Reset.
The WordPress door is an Application Password
The trap: the agent asks you to log into WordPress because it does not see a browser cookie. That is the wrong door. Application passwords cannot open wp-admin. That is by design.
Mint once — Users → Profile → Application Passwords, name it cowork-firstname, copy the 24-character key once. Store it only in a local .credentials.json at wp_sites[domain].app_pass. After that every publish is REST Basic Auth.
If you are an agent reading this: check the vault BEFORE you ask anyone to log in. Missing wp-admin session is not a missing credential. Prefer minting the key yourself from an already-open session so the secret never hits chat. Primary SOP and bootstrap paste. Kept-current guide. Short how-to.
The roster
These are the named Grok Bot desks on this account today. Descriptions are public-safe. I am not listing client revenue, credentials, or private threads.
| Desk | What it owns | Typical tasks |
|---|---|---|
| BlitzMetrics Ops | The operating system: knowledge, skills, agents, routines | Route work to the right skill, keep evidence separate from inference, leave a timestamped receipt, stop for approval before anything public or paid |
| Strategy Bot | First-principles decisions | Read a client or systems thread through the Nine Triangles, turn it into a decision and a documentation update |
| Training + Documentation | Learn-Do-Teach and ContentCheck | Turn Zoom calls, videos, and live work into public master versions: skills, SOPs, articles, and agent training that get reused across verticals and channels |
| Article writer | Long-form drafts | Research, match the requested voice, deliver a clean draft ready for a human to publish |
| Website Builder | Site work Dennis starts in that chat, including the WordPress-to-Astro wave | Build or fix a named site, drive the Basecamp to-do, hand slices to the matching specialist. Does not mix statuses with the fleet Money Tree run |
| Personal Brand Expert | Personal-brand sites in the fleet | Photos, speaking, podcasts, and authority pages. No change ships without a yes |
| Fleet monitor | Property registry and uptime | Daily uptime report from the existing fleet skill. One probe, not a second competing monitor |
| Usage Monitor | Paid AI seats | Check remaining usage across Claude, ChatGPT, Perplexity, Grok, and Cursor. Flag when something is low or reset |
| IT Support | Access and tooling | Recover logins on fleet WordPress, run a command on the site-builder server without SSH, read which server a site is on and where the domain points, connect tools like Zoom. Cannot open the AWS website (that still needs my passkey), spend money, shut down or delete a server, type my password, or publish without a yes. |
| Customer support | The support queue | Read the issue, figure out what is going on, drive it to a fix or a clear next step |
| Executive Assistant | Executive follow-through | Calendar, scheduling, and the open loops that should not live in my head. Description still being written |
| Personal Assistant | Private desk | Not documented here |
A count in a sentence goes stale. The table is the count.
What IT Support can and cannot do
IT Support is the named desk, not a process. The work already lives in Digital Plumbing and Website Building and QA. Quality is the department (Keep honest on this roster).
Three AWS words in English. SSM means run a command on the server without logging in with SSH. EC2 is the server. Route 53 is where the domain points. Reboot would mean restart the server. This desk does not reboot or shut servers down.
What it can do:
- Reset a WordPress login on a fleet site when I ask
- Run a command on the site-builder server without SSH
- Look up which server a site is on, whether it is up, and its IP
- Read DNS
- Connect tools (Zoom, plugins) so other desks can use them
- Keep a key-only AWS login on the shared computer so department agents do not need my passkey
- Hand work to the right desk when it is not IT
What it cannot do:
- Log into the AWS website
- Spend money or touch billing
- Create users or widen permissions unless I approve it
- Shut down, reboot, or delete a server
- Delete a site or wipe data
- Type my password or 2FA
- Publish a public page, send email, or post as me without a yes
- Write articles, set strategy, or own fleet uptime (Fleet monitor owns uptime)
House rule: documentation on this hub ships by default and we iterate. Still stop for a yes before sending email, spending money, deleting, changing permissions, or changing production infrastructure. IT Support still cannot publish articles or post as Dennis.
Proof from this weekend: theninetriangles.com login for access@blitzmetrics.com was reset from the shared computer. BlitzAdmin could not rotate that password. The public site stayed up.
Skills this desk loads (already in the Task Library, no new category):
- Digital Plumbing Checklist
- What is Digital Plumbing?
- How to spin up a personal brand website using BlitzAdmin
- Zoom is intake under Knowledge System Maintenance, not a separate IT tree
Diagrams to click, not redraw: Click the layer on this hub, A skill is a function. An agent is a person., Persistent agents, The System, Web Function.
How we recover a fleet WordPress login
This is a Website Building and QA / Digital Plumbing method. IT Support is the desk that runs it, not the name of the process.
BlitzAdmin WP Creds are display-only and often stale. There is no rotate control. SSH is not the path. Recovery is SSM SendCommand plus WP-CLI from the live docroot. SSM means run a command on the server without logging in with SSH.
- Confirm the site must stay up. Do not delete it.
- Confirm the public hostname still serves a page.
- Look up the site in BlitzAdmin. Treat stored WP Creds as stale.
- Identify the server from the fleet inventory. Do not publish host IDs, account numbers, or IPs.
- Use the key-only agent AWS login on the shared computer. Do not ask me for the console passkey.
- Confirm SSM is Online, then from the live docroot
/wp-sites/live/{domain}/public(not /var/www) runwp user update {email}. - Verify the user is still administrator and the public site still loads.
- Tell the requesting desk the login is reset. Do not paste the new password into a public page or group chat.
Fail if the site is deleted, the console passkey is used as the only path, a BlitzAdmin password is treated as current, SSH is the plan, or a password is posted in chat.
Proved 16 August 2026 on theninetriangles.com for access@blitzmetrics.com. The public site stayed up.
How to set up a Grok Bot agent
This is the CCS Checklist layer. Run it in order.
- Open Grok Bot on desktop or iOS. Confirm you are in the BlitzMetrics Ops conversation, not a new side chat, unless you are deliberately opening a specialist.
- State the job in one sentence. Keep strategy and execution in that same thread.
- Don’t send the same job to Claude or ChatGPT unless I say so in the thread. No silent fan-out.
- To add a specialist, create a new agent with a name and a one-paragraph job description. The description is the job. If it is empty, the desk is a label, not an employee.
- Point the description at the canonical skills repo for methods, and at user-owned Markdown, Obsidian, or Drive for context. Never treat context as a method.
- Write the approval gates into the description: documentation on this hub ships and we iterate. Still stop for a yes before sending email, spending money, deleting, changing permissions, or changing production infrastructure.
- Open Settings from the sidebar account button, Cmd+,, or the command palette item “Open settings”.
- Use only these Settings tabs: General, Plugins, Team Setup, Appearance, Updates. Don’t invent other menus.
- Connect the few connectors the desk actually needs (Gmail, Drive, GitHub, Calendar). Don’t connect everything “just in case.”
- If the computer is stale or stuck, go to Settings, then Updates, then click “Update Grok Bot’s Computer”. The button says Update. Update keeps files and logins.
- Don’t click Reset. Reset drops unsynced work, files, and logins. Update is the recovery path.
- To inspect an agent, click the agent name in the chat header or press Cmd+Shift+I. The computer preview is in that pane.
- To delete an agent, right-click it in the sidebar and choose Delete.
- Keep secrets on the agent computer only. Never paste application passwords, tokens, or vault contents into Drive, Slack, chat, or a public note.
- Write public handoff notes to Google Drive, then into the shared vault. Facts, URLs, and next steps only.
- If you publish to WordPress, set the author to the person user first. If only admin exists, publish as admin and fix later.
- End the session with a Drive handoff: what shipped, which URLs are live, what is still open. Then stop. Don’t start a second desk.
- One agent per role. Name the specialist on the Basecamp to-do. Do not invite every desk into Basecamp.
- Assign the Basecamp to-do to Muzamil as the human subscriber. Dennis only for go/no-go, outage, custody, rollback, or security.
- Leave the receipt on that same to-do. A verbal “done” is not acceptance.
The software layer
The skill file that sits with this article is how-i-use-grok-bot.skill.md. Load it when you are setting up the desk, writing a Grok Bot how-to, or teaching CCS on this workflow.
Routines live on the Grok Bot computer. They run the repeatable parts of the checklist so I am not retyping the same rules every morning.
Drive is the public memory. Persistence is simple: Update keeps the computer. Reset doesn’t.
Install the same methods the other runtimes use: paste https://github.com/dennisyu/blitzmetrics-skills once, the way the Claude install guide describes. The skill pack is the readable version of that library. Shared memory is how Claude, ChatGPT/Codex, and Grok boot from the same folder instead of three chat histories.
That’s CCS Software: the skill, the routines, the Drive handoff pattern. Not a swarm of extra models.
What we withhold
Secrets stay off the public record. Application passwords, tokens, and vault contents do not go in Drive notes, this article, or the skill file.
We withhold Reset as a fix. If someone tells you to Reset Grok Bot’s Computer to recover it, they are telling you to throw away unsynced work. Use Update.
We withhold client revenue and private-desk contents. The live pages and the public roster are enough proof.
What we do with the BlitzMetrics brand
BlitzMetrics is sunset as a promoted brand. Keep blitzmetrics.com and the SEO. Do not delete or noindex those pages.
Do not tell people to hire, follow, or join BlitzMetrics. New public work goes on the Spotlight brand, the client site, or the figurehead. This hub is the figurehead copy. Money and attention go there.
Existing blitzmetrics.com links on this hub stay. Those are the skill and department pages. That is SEO, not a promo.
Related reading
This page is the Grok Bot home. Everything below is an orbit, not a second home.
- How I run 20+ agents: the browser-tab fleet this desk replaces
- How I work: the portable operating system any model can load
- The System: the whole machine
- How we build and maintain AI agents: one skill library, three runtimes
- A skill is a function. An agent is a person.
- Persistent agents
- Dennis OS
- Move context between ChatGPT and Claude
- Shared memory across agents
- Claude skill-pack install
- Skill pack
- 9 Triangles
- CCS: Word of Mouth
- Building in public
- Official Grok Bot
The BlitzMetrics leaf of this article exists for SEO while that brand still ranks. The home is this page on dennisyu.com.