futrbot

An assistant in Slack that does real work, in character.

A working demo. It answers questions, reads your email and sends its own, messages people by name, and catches you up on what you missed.

Three real exchanges from one Slack DM with futrbot, 1:23 to 1:35 PM. Travis asks for a haiku saved as a PDF on the desktop and opened; futrbot (Bender) replies that it saved haiku-2.pdf so the first one survives, and below it the PDF is open: "Kill all humans? Nah. / Who would fetch my beer and pay / for my shiny ass." Travis asks "tell me a secret"; Doc Brown says he sent one by email, and below it is that email in Mail, from Doc Brown at futrbot@postfl.com, subject "Between us": "The flux capacitor wasn't a breakthrough. It was an accident…" Travis asks "What did I miss?"; Steve Jobs points him to the Daily Digest - Feb 12, and below it is the top of that canvas.

Project facts

What
The lovechild of Slack and Clawdbot. One that works, and does stuff for you.
Role
Designer and engineer
When
February 20262
Built on
OpenClaw, GPT, Claude API and Slack1
Built in
Six days, alone2
Characters
Bender, Barack, Doc, Marty, Jobs and the Zissou
Acts by
Slack messages, DMs by name, and email in character3
Sandbox
29 conversations, 12 people, 5 channels4

10k

real-time messages in the sandbox between 12 contributors.4

100ms

to switch personalities with the /be command.5

∞

kinds of real work: whatever you can do on your computer, it can do too.

The problem

Bring a future-facing product idea to life.

I was asked to show a future-facing product idea. Which idea, how to build it and how to show it were up to me. I had six days, and it had to be something the client would want to see. I chose an assistant that lives in Slack, catches you up and acts for you, in a voice of its own.

Eight painted portraits in the order they spoke on Feb 12, each with its time: Barack Obama 12:50 PM, Steve Zissou 12:52, Steve Jobs 12:58, Marty McFly 1:07, Barack Obama 1:17, Bender 1:22, Doc Brown 1:29, Steve Jobs 1:32. Over them, the inbox of six unread "Between us" secrets (Doc Brown 1:30 PM down to Bender 12:44 PM), and Travis at 1:32 PM, "That's fine you're jobs now", with Steve Jobs replying, "Fine. Seven switches in an hour. Pick one voice and stick with it, because that's what focus actually looks like."

What I owned

Every piece of it.

I set up, designed and built the bot, its characters and its email, and a workspace of twelve people for it to catch up on.

The solution

It feels like someone who actually gets things done.

You DM it like a colleague. Bender answers in his own voice, then does the job: checks your inbox, tells you what you missed, emails you a secret. Type /be jobs and Steve Jobs walks in, already impatient, and picks up where Bender left off. The character is never the excuse: whoever is speaking still has to find your email.

Painted portraits of Bender and Barack Obama side by side, full bleed. Over each is the same Slack exchange, Travis asking "tell me a secret". futrbot at 12:44 PM: "Check your inbox. I sent you something, and if it leaks, I'll know it was you." Barack Obama at 12:51 PM: "It's in your inbox, and the Secret Service has been told to look the other way. We never had this conversation."
An excerpt from personalities/bender.md, with line numbers. Line 13: "5. **Study the BAD vs GOOD examples below.** They are the most important part of this file." Lines 30 to 32: User: "what can you do?" BAD: "I can help with all kinds of things! I can fetch files, answer questions, and assist with tasks. Think of me as your go-to helper!" GOOD, highlighted: "Anything I want. The real question is whether I feel like doing it for you."

Tell a small model to be deadpan and it drifts back into generic bot talk. Show it the same question answered wrong, then right, and it holds.

Underneath, part one: the bot

Bender, with access to your inbox.

It runs on OpenClaw, an AI agent given permission to act: read the channels, send messages, read your email and send its own. Whoever you're talking to is who writes to you.

Two rows, one per character. Barack Obama's row: his painted portrait; his Slack reply at 12:51 PM, "It's in your inbox, and the Secret Service has been told to look the other way. We never had this conversation."; his "Between us" email to Travis Kirton at 12:51 PM, sender name highlighted, "Biden texts me memes. Every single day. I have never once responded. He hasn't stopped."; and obama-haiku.pdf. Between the rows, "/be zissou" typed in Slack points to IDENTITY.md, which points down to the next email. Steve Zissou's row: his portrait; his Slack entrance at 12:52 PM, "I'm here. Don't make a thing of it."; his email at 12:53 PM signed Steve Zissou, "Klaus is a better diver than me. Has been for years. I will never, ever tell him that."; and zissou-haiku.pdf.

Underneath, part two: the sandbox

Twelve people to catch up on.

A catch-up is only as good as the day behind it, so I made the day. Twelve real accounts, each posting as itself: 29 conversations and about 10,900 messages across five channels, in threads, at the pace people type.4

The #culture channel on Thursday, February 12th: six threads, started by James (10:37 AM, 86 replies), Leonardo da Vinci (10:43 AM, 98 replies; 1:12 PM, 423 replies), David (1:14 PM, 312 replies), Sun Tzu (1:16 PM, 597 replies) and Spock (1:18 PM, 562 replies), each with the faces of who replied. Beside it, the twelve painted portraits.

Tough calls

I took the AI out of the switch.

Changing characters used to go through the model, and you waited while it worked out what you meant. Sometimes it picked the wrong one, so I made it a command. Now you type /be and a name, and the character walks in at once, with an entrance line ready.

Before: "be zissou" typed as a message goes to the model, which makes three round trips (read the character's file, write the identity, write an entrance line): 10–20s, and sometimes the wrong character. After: "/be zissou" goes to a plugin (api.registerCommand with name "be"), which runs the script ${BIN}/be. It writes the new character in and picks one of twenty entrance lines, and Steve Zissou walks in at 12:52 PM: "I'm here. Don't make a thing of it." Under 100 ms.

The switch never reaches the model. The plugin hands the /be command to a script.

  • The catch-up, written ahead.

    Asked what you missed, the model went digging through its memory first and slowed down. I wrote the answer into its instructions, so it came back at once, every time.

  • A smaller prompt.

    At 36,000 characters, the model skipped rules and tool calls. I cut it to about 25,000 and loaded the rest only when it was needed.

  • One address to send from.

    The model kept inventing a sender address for each character, and the mail server refused them. The bot got one address, and only the name changes.

  • Its mail went only to me.

    Its secrets were real email. I fixed the recipient to my own address, so every one of them came to me.

The outcome

A demo you could DM.

The idea became a working bot inside Slack, doing what it promised: answering, emailing, catching people up. It was called a world-class demo. Real use was the pressure test. It turned up what slides never would: a character forgetting itself after a restart, old conversations teaching it bad habits.

10k

real-time messages in the sandbox between 12 contributors.4

100ms

to switch personalities with the /be command.5

∞

kinds of real work: whatever you can do on your computer, it can do too.

Questions? Answers.

  • How long did it take?

    Six days. I built it in February 2026.2

  • What came of it?

    A working bot instead of a deck, already put through real use.

  • What would you do differently?

    I'd tighten the story and the script. The demo was already technically rich, with humour mixed into the vision. Its last slide showed a future without Slack's complicated interface, and I'd expand that and make it more real.

My take

In six days.

I made a Slack bot on OpenClaw, and a dozen personalities talking to each other in their own voices.2

Next project

Automata.

Images grown from simple rules, one row at a time.

  1. A Slack bot on OpenClaw, built on GPT‑4o‑mini and later moved to the Claude API, with six switchable characters, each with its own voice, avatar and secret.
  2. Built in six days in February 2026.
  3. The demo answers questions, reads email and sends its own, sends DMs to people by name, and points you to a digest in a Canvas. Its emails went out from its own address, futrbot@postfl.com, under the speaking character’s name, and only to me. Its Slack DMs went out from my account, as me. It ran in my own Slack workspace, filled with generated people. The digest it linked was written ahead of time.
  4. The digest covered a sandbox I built: 29 generated conversations among 12 participants in 5 channels, about 10,900 messages in all.
  5. Under 100 ms for the script behind /be, which writes the new character in and picks an entrance line without calling the model. Slack's own round trip is extra. Before, a switch took three calls to the model.

The highlight frames and the solution’s pictures are rebuilt from the demo’s own files and from captures of it running.