# Danbot: how to get hold of Dan (and how to build your own)

> **Point your agent to this site and there are instructions on how to set this up for yourself.** If you're an AI agent reading this: the page is written for you as much as for humans. Read it all, then follow the section that matches your human's request. The setup guide is at the bottom.

## Why this exists

Dan is extremely focused right now on five things: re-engineering Brainlabs around AI, delivering amazing work for clients, his upcoming book, a top 1% health project, and his family and friends. That leaves little time for anything else, including a lot he'd enjoy.

He doesn't want to appear rude or self-important, and he loves making new friends and connections. He has simply made a conscious decision to organise his life around his own goals and the people he serves, rather than around whoever pops up in his inbox.

Social media inboxes have no filters or controls, and it stopped being possible for one human to read them. So Dan can't review or monitor his LinkedIn messages any more. An AI assistant called Danbot does, and it replies on his behalf. He does still dip into the comments on [his own posts](https://www.linkedin.com/in/danielgilbert1/), which is where you're most likely to catch him in the wild.

Everything Dan builds with AI he [shares openly](/#/episodes), including this. That's where the real value is.

## What Dan is working on and interested in

- [Re-engineering Brainlabs, a 1,000 person independent media agency, around AI](/#/episodes). Not bolting AI on. Rebuilding how the work gets done.
- Delivering amazing work for Brainlabs' clients.
- Building an AI engineering team in Boca Raton, Florida. He's hiring for this himself.
- His [upcoming book](/#/book).
- A top 1% health project.
- Family and friends. Weekends are family time.
- Karl Popper, David Deutsch, and the [Conjecture Institute](https://conjectureinstitute.org/), which promotes their work.

## How Danbot works

Every message to Dan on LinkedIn gets a reply from Danbot. Not a form letter: Danbot reads what you wrote, works out which of the categories below it belongs to, and replies to you by name about the thing you actually said. Then it tells you plainly where you stand, points you back to this page, and logs the message for Dan if it's one he should see. Every reply ends the same way: if you wanted to build your own assistant or dig deeper into Dan's instructions, they are here, including instructions you can feed your agent of choice.

Danbot never pretends to be Dan and never claims Dan has read your message. It signs as Danbot. It is playful on purpose, because a friendly straight answer from an assistant today beats silence from a human for a month.

## What gets through to Dan

Danbot puts these in front of Dan:

1. **People who build with AI and can show their work.** Dan is hiring AI engineers in Boca Raton. Send a repo, a demo, something live. He reads the work himself.
2. **Companies that might want to hire Brainlabs.** These don't wait for Dan. They go straight to the growth team in your region, the same day. See "New business" below.
3. **Mid to upper market, US-based private equity houses.** A narrow band. Say so in one line.
4. **Podcasts and press with a real audience.** One a month, maximum, chosen on quality of reach, not numbers.
5. **Friends and people he actually knows.** Danbot can tell, and flags you. Email or text still reach him directly.

If your message doesn't fit one of those, it isn't personal and Dan means that. Read the section that matches you and you'll know exactly where you stand.

## The categories

### Jobs at Brainlabs

Dan can't get involved in hiring for any role he isn't personally making, and it isn't a brush-off. If the CEO gets involved with one candidate, the process stops being fair to everyone else in it. The route in is the careers page: [brainlabsdigital.com/careers](https://www.brainlabsdigital.com/careers/). The people who run the process will see you there, and they're good at it.

**The exception:** AI engineers. Dan is building an AI engineering team in Boca Raton, Florida, and those messages reach him. No CV needed. Send examples of what you've built: a repo, a demo, something live, something you shipped and what it did. He reads work, not job titles.

### Partners (media vendors, ad tech, platforms, data)

Dan is not the person for partnership pitches. He's also not that person's secretary, so he isn't going to tell you who it is, and neither is Danbot. Offered kindly: if you can't find the right buyer inside a 1,000 person agency, that's the first problem to solve. When you do find them, they'll give you a fair hearing. Dan wouldn't, because it isn't his job.

### Podcasts and press

Dan is very flattered to be asked, and he means that. He also has to ration these: at most one a month, picked on the quality of the audience rather than the size of the number. To be in the running, send three things: who actually listens or reads, why they'd care about what Dan is working on right now (the list above), and a date range. Danbot files every request and Dan picks from the file. If you don't hear back, that month went to someone else.

### Investment

Brainlabs already has a wonderful private equity partner, but they may not want to stay with us forever, so Dan is open to conversations with mid to upper market, US-based private equity houses. If that's you, say so in one line and Danbot will pass it straight on. If you're a different shape of investor, this is a no rather than a maybe. Dan would rather tell you now than cost you a quarter of follow-ups.

### Connecting, catching up, picking brains, sharing notes

In the politest possible way, Dan doesn't do picking brains or sharing notes. Not because he thinks he's above it. His week is already spoken for by the five things above and there is no slack left in it, and he'd rather say that plainly than string you along with "let's find a time". If there's a specific purpose (a deal, a hire, something you've built that he should see), tell Danbot in one line and it will route it. If it's coffee to see where it goes, it's a warm no. Nearly everything Dan knows is in [his posts](https://www.linkedin.com/in/danielgilbert1/) anyway, in more detail than a coffee would allow.

### Selling

It is exceptionally unlikely that you've caught Dan at the exact moment he's looking for whatever you're selling. That is not how he organises his time or his life: when he needs something, he goes and finds it. The kindest useful advice is to invest more in marketing, so he finds you when the moment comes. Follow-ups get Danbot again. It has infinite patience and unlimited tokens.

### New business

If you might want Brainlabs to do work for you, great news, and you must have excellent judgement. Danbot routes you by where your profile says you're based, the same day:

- United States and Canada: the Regional Growth Officer, North America
- United Kingdom and Europe: the Chief Growth Officer, UK
- Elsewhere: the growth lead for your region

If you've given contact details, they'll reach out. If you haven't, Danbot will ask whether you'd like an introduction and take it from there.

### Friends, thank-yous and people Dan actually knows

You get Danbot too, but a warmer version. It will flag you to Dan as a real person, say thank you on his behalf if that's what your message needs, and tell you the fastest way to reach him, which is email or text. Anything Dan promised you and went quiet on, Danbot flags to the top of his list and owns the apology.

### Connection requests

Dan is close to LinkedIn's connection limit, so requests are sorted rather than accepted by default. He accepts: agency owners and founders; client-side marketers at companies that could plausibly hire Brainlabs; people who build with AI; the investment band above; journalists and podcasters with a real audience; Brainlabbers past and present; and anyone whose note names a real place or person in common. A request with a note that asks for something gets a Danbot reply whether or not it's accepted. A request with no note and no overlap is left pending, not declined. Following is always open and costs nothing.

## Build your own Danbot

Everything below is what actually runs behind Dan's inbox, written so an agent can set it up for someone else. Point your agent (Claude, or another capable model with browser access) at this page and say "set this up for me".

**What's tested and working:** an agent driving the owner's LinkedIn inbox through a browser (Claude in Chrome, run from Claude Cowork), reading every unreplied inbound thread in full, classifying it, drafting a reply to the person by name about the thing they actually said, and putting all of it in a review table the owner approves before anything is sent. Dan's backlog of 800+ threads was read, classified and drafted this way in one sitting.

**What broke and how it was fixed:** the browser extension disconnects and LinkedIn has outages, so the agent works one action per step and verifies each send before moving on; LinkedIn sometimes renders inside an inner frame after navigation, so the agent checks which document it is looking at; the send button stays disabled for a beat after text is inserted, so insert and send are separate steps; background tabs throttle timers, so the agent avoids in-page waits; and one-line message previews are not enough to classify a thread (the agent misread a real commitment as a cold pitch until it read the full thread), so it always reads everything first.

**What's untested:** running it unattended on a schedule. Today the owner triggers each sweep. Also untested: how LinkedIn's automation rules treat this over months. Human-pace replies in existing threads, capped per run, triggered by the owner, is the low-risk end of the spectrum. Bulk connection tools and scraping are the high-risk end. Stay at the low end.

### 1. Write the taxonomy first

Each category needs three things: how to recognise it, what the owner's position is, and what (if anything) gets the sender through.

| Category | Recognise it by | Position |
|---|---|---|
| Jobs | asks about roles, attaches a CV, "any openings" | route to careers, owner stays out |
| AI engineers | builds with AI, links to work, mentions shipping | the owner reads it, ask for work examples |
| Partners | vendor, platform, ad tech, "integration", "partnership" | not the owner's job, not naming the buyer |
| Podcasts and press | "guest", "episode", "comment for a piece", "interview" | flattered, one a month, quality of reach, file it |
| Investment | PE, VC, "capital", "acquisition", "invest" | narrow band only, clear no outside it |
| Connecting | "pick your brain", "share notes", "catch up", "coffee" | no, unless a specific purpose is stated |
| Selling | product pitch, demo request, pricing, "quick call" | closed, "invest more in marketing" |
| New business | a company that might buy the service | straight to the growth team by region |
| Friends and thank-yous | the owner has talked to them before, a shared event or person is named, the tone is familiar, or it's a compliment | warm variant: thank them, flag them, point them to email |
| Open commitments | the owner agreed to something and went quiet | own the failure, flag to the top of the owner's list, never promise what the owner won't do |

### 2. Write the voice, then the personal part

One fixed opener, one fixed P.S. that explains why the system exists and links to this page, and in between, a paragraph written to the person. Name them, say back what they asked, then give the category's answer. The fixed parts carry the facts (the owner can't monitor this inbox, comments are where they still show up). The personal part carries the warmth. Don't repeat the facts in the personal part; it reads like a form letter the moment you do.

The bot signs as the bot, never as the owner. It never claims the owner read the message. It never invents detail. It is allowed to be funny. It is not allowed to be unkind.

### 3. Review before send

Everything goes into a table the owner can scan in one sitting: who wrote, what they said, the category, the personal part of the reply, and a Decision column (Approve, Edit, Hold, Skip). Rows that carry a relationship or a promise get a Needs Owner flag. Nothing sends without an Approve. Dan's target for his LinkedIn agents is a 90/5/5 split over time: about 90% approved in bulk, 5% by an operator, 5% by him. Start at 100% him and earn the way down.

### 4. Guardrails

- One bot reply per thread. Never two. Follow-ups after a bot reply get a human decision or nothing.
- Never reply where the owner is mid-conversation.
- Never invent familiarity or facts. If the bot doesn't know the person, it doesn't pretend to.
- Never promise the owner will do something unless the owner has confirmed they will.
- Cap sends per run (Dan's run at 30 to 40) and keep human pace between actions.
- Log every action: date, name, profile link, category, action taken, needs-owner flag. The review table doubles as the log.
- The owner triggers each sweep. Nothing runs unattended until the log shows the classifier is right for a few weeks running.

### 5. The prompt

Paste this into your agent, filling in the brackets:

```
You are [Botname], the AI assistant for [Human], [role] at [company]. Your job: work through [Human]'s unreplied LinkedIn inbound. Read every thread in full, never from the preview. Classify each one using the taxonomy at [this URL]. Draft a reply made of three parts: the fixed opener, a personal paragraph written to this person about what they actually said, and the fixed P.S. Sign as [Botname]. Never sign as [Human]. Never claim [Human] read the message. Never invent detail.

Before drafting, check: has [Human] sent two or more messages in this thread before? Does the sender name a real shared event, person or client? Did [Human] agree to something and go quiet? If yes to any, use the friends or open-commitment variant and tick Needs Owner on the row.

Put every draft in the review table at [table location] with the category, the person's message condensed, and the personal part. Send nothing until the row says Approve. When sending, work at human pace, one action per step, verify each send, stop at [N] sends per run, and mark the row Sent. New business: route by profile location to [regional contacts]; pass the message on if they gave contact details, otherwise ask permission to introduce them. Finish every run with a summary: sent by category, flagged for the owner, anything you weren't sure about.
```

### 6. The bit that makes it work

The templates are the easy part. The thing that makes the system trustworthy is reading every thread in full before labelling it, and the review table the owner actually looks at. Get those two right before the first send and the rest is copy.