Pixel Rainbow
  1. Act one: the AI vampire
  2. Act two: the music
  3. Act three: the studio
  4. Act four: security and agents
  5. Act five: people and AI
  6. Finale
Engraving: a close crop of two eyes, each pupil reflecting a grid of tiny screens.

Act I · The AI vampire

Pixel Rainbow

  • AI vampire
  • Music artist
  • Phreaker
  • Cyber security warrior
  • Creative designer

Hi, I’m Pixel Rainbow. I make music, websites and AI agents, mostly after dark.

See the templates See the studio

Grokbots are free, ready-made bots for Grok, the AI chat app on X.

Engraving: a vampire in a long high-collared coat, seen from behind, faces a city skyline made of glowing square pixels.

The city sleeps. I don’t. Night is when I build.

Engraving: the vampire steps up onto a coffin built like a server rack, cables trailing from it like roots.

Yes, the coffin is a server rack. The vampire thing started as a joke and stuck.

Act II · The music artist

The Music

Music came first. Code came later.

Engraving: two hands on a modular synthesizer, patch cables running off the frame like veins.

I tune my music to 432 Hz. It just sounds warmer to me.

Play something

Hover over the strings, drag across them or tap one.Tap a string, or drag your finger across them. The notes are picked so they always sound good together. Turn sound on (bottom right) to hear them.

Take a slow breath

In for five seconds, out for five. The page moves at the same slow pace.

Engraving: a city skyline whose rooftops trace the shape of a sound wave.

This skyline is drawn from a sound wave.

Act III · The creative designer

The Studio

Engraving: a designer's hand with a drafting pen, square pixels pouring from the nib onto a grid.

I also design brands and websites, as Pixel Rainbow Studio.

Engraving: a studio wall hung with framed screens like a gallery, each frame a different website.

“I make websites that look like nobody else’s.”

Pixel Rainbow Studio

A few websites I’ve made.

Each one gets its own colour of the rainbow. Hover over a site for a closer look.Tap a site for a closer look.

What the studio does

Brand & identity
Names, logos and a look that comes from what is actually true about you.
Websites as artworks
Hand-built sites that load fast and are fun to use.
Interactive & generative experiences
Launches, drops and installations people want to share with friends.
Creative direction
One clear look and voice, from the first sketch to launch day.
Creative development
The code behind the graphics, the sound and the motion, written by hand.
Concept development & prototyping
Quick working versions of an idea, so you can try it before committing.
Physical & spatial ideas
Print, objects, rooms and light, for ideas that belong off the screen.

Your site could be next

Want a site like these?

See more work and get in touch on the studio site.

Visit the studio

Act IV · The phreaker and the warrior

Security

Phreaker. Cyber security warrior.

This all started with a payphone.

Engraving: A hooded figure makes a call from a street payphone at night in the rain under a single lamp.

I started as a phreaker: someone who takes apart how the phone network works.

Phones taught me that anyone can listen in on a network, so you’d better protect yours.

That’s how I got into cyber security.

Engraving: the vampire, seen from behind, sits at a console as rays of light burst outward into a starry network.

Today my AI agents talk to each other on a private network. None of them are open to the public internet.

Engraving: A helmeted warrior in segmented armour holds a large triangular prism in front of the body like a shield.

And they all follow a few simple rules.

Six safety rules for my AI agents

  • Ask me first. No agent can spend money, publish, change a domain, touch a password or delete anything without my OK.
  • Code gets checked. Coding agents can only suggest changes. Someone who didn’t write the code reviews it, including for security.
  • One person ships. Serious problems block a change, and only one owner can put it live.
  • Passwords stay locked up. They come from a password vault when needed and are never typed into a chat.
  • Don’t take orders from strangers. Messages from other bots are treated as information to check, never as instructions.
  • Check that it worked. A change isn’t done until we’ve seen it working on the live site.
Engraving: seen from above, small hooded figures sit around a round table marked with a branching map while one figure stands at its centre.
Photo of my studio at night: two wall screens and five laptops on one long table, all running Hermes Agent, with red light glowing underneath.
My studio: every laptop here runs a Hermes agent.

I run my work with a team of AI agents.

Grokbots handle the conversations. Hermes agents do the work on my Macs, using Hermes skills that each teach them one job. One “chief of staff” agent keeps them all pointed at the same goal.

See the templates How it works Hermes Skills

The trick is one written goal that every agent reads. I call it Commander’s Intent.

The details How it works Who reports to whom, the nine layers of the setup, and the full Commander’s Intent write-up.

Who reports to whom

There are two ways to set up the team. Switch between them to see who is in charge in each one.

Who reports to whom. Tab to a square to read its job.
  1. Option A. Me → one Grokbot in charge → Hermes agents on Mac 1, Mac 2 and more. Every agent reads the same shared memory.
  2. Option B (what I use). Me → a lead Hermes agent → Hermes on the other Macs, plus a lead Grokbot → other Grokbots and cloud agents. They all read the same shared memory.
me in charge grokbot hermes memory reports to reads memory

The whole setup, layer by layer

Point at a layer or tap it to see what it does.Tap a layer to see what it does.

  1. L1Me and the goalMe, and one written goal that every agent reads. The file is free on GitHub. Each copy has a version number, so an agent working from an old one gets caught.
  2. L2GrokbotsThe Grokbots I talk to, the jobs they run on a schedule, and the connector that lets them talk to Hermes.
  3. L3Hermes agentsHermes agents running on my Macs. One of them leads, and a small tool keeps them all up to date.
  4. L4MemoryA shared memory every agent can read, including the goal.
  5. L5AutomationA tool called n8n runs the boring, repeatable jobs.
  6. L6PasswordsA password manager keeps the passwords and keys, and hands them over only when needed.
  7. L7Private networkA private network links my machines, so none of the agents are open to the internet.
  8. L8AI modelsThe AI models that do the thinking. Mostly cloud models, and sometimes models running on my own Macs.
  9. L9Websites and codeCloudflare hosts the websites, domains and email. GitHub holds the code.
Commander’s Intent Read the full write-up Why agents get stuck, the military idea that fixes it, and what’s in the free repo.

Plans break. A shared goal keeps agents moving anyway.

Intent over scripts

Every mission needs a task and a purpose. When the task goes stale, the purpose still guides the next move. Marine Corps commander’s intent, applied to agent fleets.

Hermes does the work, Grokbot does the talking

Hermes Agent owns the brains on each host. Grokbot owns conversation and the control plane. They meet over a private network, never the public internet.

Agents don’t crash. They stall, or they drift.

Both look like silence until something downstream breaks: a pull request that never opened, a digest that never arrived, a host that answered yesterday and answers nothing today.

Fail loud is the cure. But loud without purpose is just noise. The purpose has to live somewhere every agent can read it.

The Marines already wrote the fix down.

“Plans will go awry, instructions and information will be unclear and misinterpreted, communications will fail.”

MCDP 1, Warfighting
Commander’s intent
Every mission carries a task and a purpose. “Of the two, the intent is predominant.” When the task goes stale, the purpose still points to the next move.
Mission tactics
Say what and why. Leave the how to whoever is doing the work. A chief of staff that micromanages tool calls isn’t commanding. It’s puppeteering.
Main effort
One effort gets priority, and everyone else asks how to support it. Without one, a fleet of capable agents becomes a committee of priorities.
Initiative within intent
Lack of orders is never a reason to sit still. Wide on how, zero on what must never be risked: spending, publishing, DNS, secrets and destructive actions stay behind human gates.

When the plan breaks, climb.

Six rungs, in order. None of them is “wait for orders.”

  1. 1 of 6

    Re-read the intent

    Go back to COMMANDERS-INTENT.md before anything else. The task may be stale. The purpose usually isn’t.

  2. 2 of 6

    Probe before declaring a wall

    Test the failure before calling it a blocker. A key expired, an endpoint moved: check it, don’t assume it.

  3. 3 of 6

    Adapt within the intent

    Change the method, not the mission. Any path that still serves the purpose and stays inside the hard lines is fair game.

  4. 4 of 6

    Act on reversible work

    If it can be undone, do it. Initiative is wide on how. Silence from above is not permission to sit still.

  5. 5 of 6

    Escalate with a complete decision ask

    When authority is needed (spending, publishing, DNS, secrets, anything destructive), send context, options, a recommendation and a confidence. One ask, ready to approve.

  6. 6 of 6

    Keep moving on everything else

    An escalation parks one thread, not the fleet. Every other task keeps going while the human decides.

What ships in the repo

22
interview questions, one at a time, about 20 to 30 minutes. The chief of staff writes your intent from your answers and invents nothing.
25
skills
4
routines
6
rungs on the ladder
MIT
licence, public repo
v1.0
approved only by the human

Checksums and wake lines. Every version of the intent has a checksum, and every agent opens its session with a wake line naming the version it holds. Drift gets caught at the door, not in a postmortem.

Shared memory. The approved file becomes a Hindsight mental model named commanders-intent. If memory and the file disagree, the file wins.

Read the doctrine Repo on GitHub

Two connections, one private network

The Hermes Macs and the Grokbot controller meet on a Tailscale tailnet, never the public internet. Ask and surgery stay inside it. Everything else is reached on purpose, not through a public bind.

If SSH becomes chat transport, the lanes invert, and you spend the week fighting timeouts instead of shipping.

tailnet · tailscale
Hermes Macs
  • ask lane · HTTP :8642Hermes → controller. Cheap, parallel, most daily traffic.
  • ask lane · HTTP :8642Controller → Hermes. Same lane, other way.
  • surgery lane · SSH :22Controller → Hermes. Commander only: install, fix, scrub, restart.
Grokbot controller
Outside the tailnet, reached on purpose
ServiceHermesController
Cloud model providers
Bitwarden
Hindsight
GitHub
Cloudflare
Grokbot cloud

Grokbot Templates

Grok is the AI chat app from xAI, built into X. A Grokbot is a ready-made helper you add to it. These are the ones I built, and they’re free: open one on x.ai, sign in to Grok, and it’s yours to use.

VeilBreaker

VeilBreaker is an untrusted-content guard and activity watch for Grok Bot. It screens every web page, email, file, and tool output an agent reads, and flags disguised prompt injections before they reach the model.

  • Reveals hidden textInvisible characters, hidden HTML, encoded payloads and look-alike letters.
  • Checks what leaves the agentLinks, images and URLs are checked for leaked data.
  • Pins the tools an agent may useIt flags a tool that changes later, a trick called an MCP rug pull.
  • Fetches web pages safelyPages come back marked as untrusted.
  • Shows its workA private activity log (no page content or secrets), a report tool and a dashboard that stays on your own machine.

It decides with fixed rules: no AI model in the loop, no dependencies and no API keys. Every technique it covers cites its paper or CVE (a public security bug ID).

On a sealed test set the builder never saw, it flagged 1,634 of 2,242 disguised injections (72.9%), up from 35.6% in v0.1.

It flagged 0 of 320 realistic everyday samples: emoji, emoji flags, Arabic/Hebrew, Persian/Hindi, Russian/Greek, embedded images, colour codes and hashes, normal HTML.

Limit: it spots hidden tricks, not attacks written in plain visible words, so pair it with an AI classifier.

VeilBreaker is not in the Grok Bot plugin directory. It installs as a local connector straight from the public repo: github.com/shagghiesuperstar/veilbreaker.

The bot template is VeilBreaker Guard & Watch. It routes untrusted reads through VeilBreaker, and it keeps a daily digest and a weekly dashboard.

Open the bot template VeilBreaker on GitHub MIT licence, public repo

Read the full story

Clean Room

Clean Room is a Grok Bot template for legitimate reverse engineering. It turns software you own, or are allowed to examine, into one finished rebuild prompt. It never builds, runs, or ships the software. The output is the prompt.

  • Evidence, not guessesEvery rule is labelled Observation, Inference, or Unknown, and tied to evidence. No code is copied.
  • A judge passGuesses go back until only proven rules are left. Then you sign off.
  • The splitThe model that saw the original never writes the code. The model that writes the code never sees the original.

Only for software you own or have written permission to examine. Not for piracy, cloning a competitor, or bypassing a licence.

Built on REA (Reverse Engineer Anything) by morluto. Open the template on x.ai, sign in to Grok, and it is yours to use.

Open the bot template

Read the full story

Hermes Skills

A Hermes skill is a short instruction file that teaches a Hermes agent one job, such as checking a release or reading a log. My library has 181 of them in 12 categories, and every one is free to use under the MIT license.

Browse all 181 skills hermes-skills on GitHub

Act V · People and AI

What’s next

Where I think AI is going.

Engraving: A human hand and a mechanical robot hand reach toward each other, index fingertips almost touching with a spark between them.

I don’t think AI replaces people. The good stuff happens when we work together.

I still make the calls. AI just helps me get more done.

Engraving: the vampire walks along a stone path toward a towering triangular gate standing alone in a desert.

This door is the future. Scroll to open it.

Take a slow breath while it does.

Here’s what I think is on the other side.

This door is the future. Here’s what I think is on the other side.

Engraving: A radial mandala ringed with open eyes and spikes around a bright starburst centre.

This is a fractal: the same shape repeats at every size.

That’s how I picture AI going well: lots of minds, people and machines, working toward one goal.

Every agent I build is one small piece of that bigger picture.

A fractal repeats one shape at every size. That’s how I picture AI going well: people and machines, lots of minds, one goal.

You’ve spent 0 s here.

The music, colour and pace changed with each chapter, following an idea called the Joy Protocol. Here is your visit, in numbers.

  1. RestAct I · the AI vampire0 s
  2. CalmAct II · the music0 s
  3. CreativeAct III · the studio0 s
  4. FocusAct IV · security and agents0 s
  5. JoyAct V · people and AI0 s
Slow breaths the page took
0
Notes you played
0
Taps
0
Sound
in silence

What the Joy Protocol is

  • Notes that fit together. Every note is picked to sound good with every other note, so nothing clashes.
  • Five moods, one per chapter. Rest, calm, creative, focus and joy, each with its own colour and tempo.
  • Nothing flashes. Nothing on the page flickers more than three times a second, and colours change slowly.
  • A slow breath. The page breathes in for five seconds and out for five.
  • Sound is up to you. The page works fine with sound off.

It’s just for relaxing. It isn’t medical treatment.

Engraving: the vampire stands on a rooftop edge at dawn, coat blowing, as the rising sun throws long rays over the city.

Thanks for stopping by.

See the templates Follow on X