---
title: "Slack Company Brain With n8n and Claude: How the Design Works"
date: "2026-10-28"
video_id: "7701789388740644110"
tiktok_url: "https://www.tiktok.com/@pro.glitch/video/7701789388740644110"
creator: "@pro.glitch"
category: "AI"
tags: ["n8n", "slack bot", "claude sonnet", "company knowledge base", "ai agent"]
excerpt: "How a Slack company brain built on n8n, Claude Sonnet and Google Docs files approved facts and answers from quotes, and what parts the demo never shows working."
---
A "company brain" in Slack is a bot that answers team questions from a set of shared documents and refuses when the answer is not there. A TikTok from @pro.glitch (Professor Glitch) shows one built with Slack, n8n, Claude Sonnet and Google Drive. The design is sensible and the frames match the description, but the company in the demo is fictional, and the key safety check, matching quotes against real pages, is described rather than shown.

The saved copy of this video had no audio track or transcript, so this breakdown relies on the frames and the post description.

## What tools does the Slack company brain use?

A diagram labeled "1 FOUR TOOLS" assigns one job to each:

- **Slack**: "where the team talks"
- **n8n**: "the wiring"
- **Claude**: "the AI · Sonnet"
- **Google Drive**: "where the brain lives"

The description says the system runs "three AI workflows, using n8n's built-in AI agent with Claude Sonnet." The frames show one of them, an n8n AI Agent node titled "File it into the brain (AI Agent)". Its system prompt begins "You file new facts into Harborline's company brain: Google Docs pages." Max Iterations is set to 10 and "Return Intermediate Steps" is on. The other two workflows, filing from an Inbox folder and a Monday report, appear only in the description. Neither the video nor the description says which Sonnet version is used.

n8n's [Anthropic Chat Model node](https://docs.n8n.io/integrations/builtin/cluster-nodes/sub-nodes/n8n-nodes-langchain.lmchatanthropic/) is the sub-node for using Claude chat models with conversational agents, which fits this setup.

## How do facts get into the brain?

According to the description:

1. A manager adds a ✅ reaction to a Slack message.
2. n8n checks an approvers sheet.
3. The original message is saved to a Raw folder.
4. The agent updates the relevant Google Docs pages.

The agent's prompt says "Every item you get has already passed the appro..." before the frame cuts off. That places the approval check in n8n nodes ahead of the agent, so the model does not decide who is allowed to teach it.

## What Slack permissions does the bot need?

The description says "the bot has four permissions and can't read private channels or DMs." The video's scope list is not visible in the frames reviewed. Slack's [reaction_added documentation](https://docs.slack.dev/reference/events/reaction_added) lists `reactions:read` as the required scope for that event and says it is delivered for content the user can see. Whether the bot is blind to private channels depends on which channel-history scopes you grant, so check that in your own app configuration.

## How does it avoid making things up?

The description says the assistant must "bring a quote copied word for word from a page it opened, and n8n checks that the quote is really on that page before the answer goes out." Links "come from the map, never from the AI."

Two answer paths appear on screen:

- **Answered.** Asked "Who is building the Northgate site, and can they launch it on a Friday?", Brain names Jordan Ellis, says the rules forbid Friday launches, gives a launch date of Thursday 2026-11-19, and cites two pages: "Northgate Property Management" and "How We Do Things".
- **Refused.** Asked about the refund policy for a project cancelled halfway, Brain says "That's not in the brain yet, so I won't guess. I've logged it to Gaps for Priya."

The frames do not show the quote-matching node, and they do not show a case where a bad quote was blocked. A card captioned "Claims vs real edits" shows a "CLAUDE SAYS" box reading "Changed the price on Shopify store", which suggests the workflow compares what the model claims it changed with the real document edits. The implementation is not shown. Treat the no-invented-quotes guarantee as a design goal until you can test it.

## Is the demo data real?

No. The description says Harborline is a company "we set up for this." Jordan Ellis, Northgate, the launch date (three weeks after the video was posted) and the description's "$24,000, was $22,000 until 2026-10-27" example are all demo data, not customer results.

## How far does this design scale?

The brain uses an Index page that lists every page. The agent reads the index and then opens only the relevant pages. The description admits the limit: "Brain reads the map every time. Past a few hundred pages, you'd add a search step." That threshold is the author's estimate.

The description also lists a Monday report covering conflicting facts, sections unreviewed for 90 days and open questions. A loop diagram shows "Monday report" next to n8n, but no report appears on screen.

## Can you get the template?

Not from the video. The description says "I'm dropping this system inside my community soon" and points to "the link in my profile". There is no URL in the description, so the build is not available beyond what the frames show.

## Bottom line

The design is worth copying: approvals checked outside the model, answers that must quote a real page, and an explicit "I won't guess" path that logs gaps for a person. The demo runs on a made-up agency, and the quote check, Slack permission set and Monday report are claims in the description that the frames do not show working. The author's own advice is to start small.

## Sources

- [reaction_added event, Slack developer docs](https://docs.slack.dev/reference/events/reaction_added): required `reactions:read` scope and delivery rules.
- [Anthropic Chat Model node, n8n docs](https://docs.n8n.io/integrations/builtin/cluster-nodes/sub-nodes/n8n-nodes-langchain.lmchatanthropic/): using Claude chat models with n8n agents.
- [TikTok post by @pro.glitch](https://www.tiktok.com/@pro.glitch/video/7701789388740644110): the video and description reviewed here.


## Video description and verification notes

**The file is a 408.6 s vertical clip (1080x1920, 9:16), MP4/MOV container (isom), one HEVC (H.265 Main) video stream at 30 fps, ffprobe stream bitrate 353,509 bps, container bitrate 295,921 bps, 15,114,207 bytes, metadata format id bytevc1_1080p_359759-1 (metadata.json lists the codec as h265 and fps as null), and ffprobe shows no audio stream in this download even though metadata.json lists aac. A missing audio stream in the file does not prove the original TikTok is silent. It was posted at epoch 1793217600, which is 2026-10-28 20:00:00 UTC, by @pro.glitch (nickname "Professor Glitch", uploader id 7505864130625733662), with audio credited as "original sound - Professor Glitch". The description is roughly 3,000 characters. Engagement at archive capture (2026-10-29 20:03:31 UTC) was 2,124 views, 73 likes, 0 comments, 12 reposts and 61 saves. No transcript.txt, transcript.json or transcript.srt could be read from the staged artifact (all three reads returned "file does not exist"), so I have 0 transcript words and 0 segments, and I could not run the transcript-end-versus-audio-duration check. Nothing below is attributed to the narrator's speech. I read 6 frames: frame_001, frame_020, frame_050, frame_100, frame_150 and frame_200, and I did not read the frames between them, so the walkthrough below is partial. Frame_001 shows a Slack thread titled "Thread" in #ask-brain with the caption "A company brain inside Slack" and a tool-logo badge (Slack, n8n, Claude, Google Drive). A blurred user asks "@Brain Who is building the Northgate site, and can they launch it on a Friday?" and the app Brain replies "Jordan Ellis is building the Northgate site. No, they can't launch it on a Friday, because the rules say no launches on Fridays. Northgate's launch is set for Thursday 2026-11-19, which is not a Friday." with the source line "Northgate Property Management · How We Do Things". Frame_020 is a diagram labeled "1 FOUR TOOLS" with Slack ("where the team talks"), n8n ("the wiring"), Claude ("the AI · Sonnet") and Google Drive ("where the brain lives"), captioned "Four tools one job each". Frame_050 is labeled "3 THE WIRING" and shows an n8n AI Agent node titled "File it into the brain (AI Agent)" whose system prompt begins "You file new facts into Harborline's company brain: Google Docs pages. Today is" followed by an expression using the America/New_York zone, with Max Iterations set to 10 and "Return Intermediate Steps" switched on; the caption reads "n8n AI agent + Claude". Frame_100 is labeled "HOW IT LEARNS", with a card "Two more guards" and a "CLAUDE SAYS" box reading "Changed the price on Shopify store", captioned "Claims vs real edits". Frame_150 shows a second Slack thread where a user asks "@Brain what's our refund policy if a client cancels a website project halfway through?" and Brain answers "That's not in the brain yet, so I won't guess. I've logged it to Gaps for Priya.", captioned "I won't guess". Frame_200 is a diagram labeled "THE LOOP" with "teaches", "asks" and "Monday report" tags and the caption "Start small". The spoken argument is unavailable, so the argument I can report is the one in the description: filing facts from approved ✅ reactions into linked Google Docs, answering only from quoted page text, and sending a weekly report.**

## The architecture claim: the frames match the description

The description says the system has "three AI workflows, using n8n's built-in AI agent with Claude Sonnet." Frame_020 shows exactly the four-tool split the description lists, with Claude labeled "the AI · Sonnet", and frame_050 shows an n8n AI Agent node with a system prompt that names Harborline. That is direct visual support for one workflow (the filing agent) and for the choice of Claude Sonnet as the model family. I did not read any frame showing the other two workflows (the Inbox-folder filing and the Monday report) as n8n canvases, so those two are supported only by the description text.

The n8n documentation page I fetched describes the Anthropic Chat Model node as the way to use the Claude family of chat models with conversational agents, and it sits under n8n's sub-nodes. The page did not name the AI Agent node explicitly, so the pairing is consistent with the video but I am inferring it from context. The description does not say which Sonnet version is used and frame_050 does not show one. I did not fetch Anthropic's model docs, so I make no statement about which Sonnet versions exist or what they cost.

## The "✅ from a manager" approval flow is plausible, with one permission detail unverified

The description says a manager adds a ✅ to a Slack message, n8n checks an approvers sheet, and the original message is saved to a Raw folder before the agent updates pages. Frame_050 supports the gating idea: the visible prompt text says "Every item you get has already passed the appro..." followed by a line about checking "(only an active approver's ✅ or an Inb..." and the text is cut off at the frame edge, so I can only quote the beginning. This tells us the approval check happens upstream of the agent, in n8n nodes, rather than being left to the model's judgment.

Slack's reaction_added event documentation says the required scope is reactions:read, and that the event is delivered for users who can see the content that was reacted to. The page did not address private channels directly. The description claims "the bot has four permissions and can't read private channels or DMs." I could not confirm the count of four, because the video's scope list is not in any frame I read and the Slack page lists only reactions:read for this one event. Whether the bot is blind to private channels depends on which channel-history scopes were granted, and that is a configuration fact I could not check. I classify the "four permissions" and "private channels stay private" statements as unverified.

## The "won't guess" behavior is shown once, on one question

The description says the assistant must "bring a quote copied word for word from a page it opened, and n8n checks that the quote is really on that page before the answer goes out," and that links "come from the map, never from the AI." Frame_001 shows a successful answer with a source line naming two pages, "Northgate Property Management" and "How We Do Things", which is the output shape the description promises. Frame_150 shows the refusal path: Brain says "That's not in the brain yet, so I won't guess" and states it logged the question "to Gaps for Priya." Both match the description's items 5 and 6.

What the frames do not show is the verification step itself. In the frames I read, no n8n node does the quote-matching and no failing case appears where the model quoted text that was absent from a page and was blocked. Frame_100 hints at a related check: a "CLAUDE SAYS" card reading "Changed the price on Shopify store" under a "Claims vs real edits" caption, which suggests the workflow compares what the model says it did with what actually changed in the documents. The card is visible only as its first state, and I did not read adjacent frames, so I cannot say how that comparison is implemented or whether it caught anything. The guarantee that answers cannot contain invented quotes is therefore a design claim, not a demonstrated result.

One reading detail in frame_001 deserves attention: the answer says Northgate's launch is "set for Thursday 2026-11-19." The video was posted 2026-10-28, so that date is three weeks after posting. This is consistent with a fictional agency, and the description confirms Harborline is a company "we set up for this." That makes every figure on screen, including the "$24,000, was $22,000 until 2026-10-27" example from the description and the "Shopify" price card in frame_100, demo data and not a customer result.

## The scaling limit the description admits, and what the video does not claim

The description states: "Brain reads the map every time. Past a few hundred pages, you'd add a search step." This is a candid limit and it is consistent with the architecture: an Index page lists every page, the agent reads it, and then opens only the relevant pages. Nothing in the frames contradicts it, and I found no source to test the "few hundred pages" threshold. It is the author's estimate and I mark it unverified.

The description also lists a Monday report covering conflicting facts, sections unreviewed for 90 days, and open questions, and says conflicts are checked "word for word on their pages." Frame_200 shows "Monday report" attached to n8n in the loop diagram, but I read no frame displaying a report. The 90-day staleness window and the contradiction check are description-only claims.

## Where the video sends viewers

The description says "I'm dropping this system inside my community soon" and "if you want to learn AI and automation from scratch, check the link in my profile." The description contains no URL, and no links directory was available in the staged artifact (links/links.json could not be read). Following the profile link would require visiting tiktok.com, which I did not do because its robots.txt disallows Claude's agents. So I could not determine whether the destination is a free resource, a lead form or a paid community. The template itself is not released: the author says the system arrives "soon" inside the community, so a viewer cannot rebuild the exact workflow from the video alone beyond what the frames reveal.

## Key Takeaways

- **Verified:** Frame_020 shows the four-tool design (Slack, n8n, Claude Sonnet, Google Drive) matching the description, and frame_050 shows an n8n AI Agent node, "File it into the brain (AI Agent)", whose prompt names Harborline.
- **Verified:** Frames 001 and 150 show the two answer paths the description promises: a sourced answer naming two pages, and a refusal ("That's not in the brain yet, so I won't guess") that logs the question to Gaps.
- **Clarification:** Harborline is a staged 14-person agency "we set up for this" per the description, so the Jordan Ellis, Northgate and $24,000 figures are demo data and not customer results.
- **Unverified:** The word-for-word quote check, the "four permissions" Slack scope set, the 90-day staleness rule and the "few hundred pages" scaling limit are stated in the description but not shown in any frame I read, and the Slack reaction_added page lists only reactions:read for that event.
- **Context:** No transcript or audio was available (the downloaded file has no audio stream), I read only 6 frames, and the profile link behind the community offer could not be followed because it sits on tiktok.com.

## Resources

- [reaction_added event, Slack developer docs](https://docs.slack.dev/reference/events/reaction_added) establishes that the event requires the reactions:read scope and is delivered for content the user can see, without addressing private channels.
- [Anthropic Chat Model node, n8n docs](https://docs.n8n.io/integrations/builtin/cluster-nodes/sub-nodes/n8n-nodes-langchain.lmchatanthropic/) establishes that n8n has a sub-node for using Claude chat models with conversational agents.
