How To Build An AI Second Brain (And Why You Need One Now)
Video by AI News & Strategy Daily | Nate B Jones on YouTube · 30 min · uploaded January 2026
Notes and analysis by BrainTube, written from the video. Not affiliated with AI News & Strategy Daily | Nate B Jones. Watch on YouTube
In short
Nate B Jones lays out a no-code capture loop: dump every thought into one private Slack channel, let Zapier send it to Claude or ChatGPT for classification, file it into four Notion databases, and get a daily and weekly digest back in Slack. The design principles behind it, especially the confidence threshold, the audit log and separating memory from compute and interface, are sound and widely useful. It is a task and relationship tracker more than a knowledge base, and a few pieces, like the one-reply fix button, are harder to build than the video suggests.
The argument
Jones argues that notes systems fail because they ask you to organise at the worst moment, while you are walking into a meeting or about to sleep, so most people stop trusting the pile and stop capturing. What changed by 2026, in his view, is that AI can run a loop on your behalf: classify a thought, route it, extract details, write it to a structured store and push the important parts back to you on a schedule. The human's job shrinks to one behaviour, capture. He presents the build in eight plain-language building blocks and twelve engineering principles, so non-engineers can reuse the pattern elsewhere, and he deliberately picks Slack, Notion, Zapier and a chat model because most people already have access to them.
Key ideas, with timestamps
- 3:31
Systems fail at capture time
He says traditional storage tools work for roughly one in 20 people he talks to, usually the already-organised ones. The wall is not motivation but timing: deciding on a folder, tag or name when you just want the thought out of your head. Notes go into Apple Notes to be sorted later, later never comes, and trust collapses.
- 5:31
The four-tool stack
Slack is the capture point (one private channel he calls SB Inbox, one message per thought). Notion holds four databases, people, projects, ideas and admin, plus a fifth called inbox log. Zapier wires them together, with Make named as a cheaper option, and Claude or ChatGPT does the classification using a prompt that returns JSON.
- 13:01
The receipt and the bouncer
Every capture is written to the inbox log with the original text, where it was filed, its new name and a confidence score from the model. At 13:31 he adds the guardrail: below a threshold he suggests as 0.6, nothing is filed, the item is marked needs review and Slack asks you to repost with a prefix such as person or project. This is what stops the store becoming a junk drawer.
- 14:30
Digests that come to you
A scheduled zap queries Notion each morning and sends a short Slack DM: top three actions, one thing you may be avoiding, and one small win. A Sunday version reviews the past seven days and names open loops, three actions for next week and a recurring theme. At 20:31 he caps them at 150 and 250 words so they fit on a phone.
- 16:02
The fix button
After each filing, the bot replies in the Slack thread with what it did and its confidence, for example a project called website relaunch at 0.87. Replying "fix:" plus a correction is meant to move the record. His point is that people only correct systems when correction costs one step.
- 17:31
Separate memory, compute and interface
He calls this the most important principle. Notion is memory, Zapier and the model are compute, Slack is the interface, so you can swap Slack for Teams, Claude for ChatGPT or Notion for Airtable without rebuilding. At 18:31 he explains why he does not start non-engineers on Obsidian: local markdown is great for ownership, but cloud automations writing into local files adds sync plumbing.
- 19:00
Prompts as contracts
A reliable agent prompt has a fixed input, a fixed output, valid values and rules for ambiguous cases, and returns JSON only. At 21:00 he adds that project entries need a next action field, because a vague intention like working on the website cannot be executed, while emailing a named person to confirm a deadline can.
- 23:00
Design for restart and keep it small
The operating rule after a missed week is not to catch up but to do a ten-minute brain dump and resume. He also argues for four buckets and very few fields, adding modules such as voice capture or calendar prep only after the core loop is trusted. At 26:30 he totals the build at three zaps, four databases plus a log and one channel, about an hour or 90 minutes.
What it gets right
One capture point with zero decisions is the single best predictor of whether a notes habit survives, and the video makes the case clearly. Moving classification off the human and onto a model is a reasonable use of AI in 2026, since sorting a short message into four buckets is an easy task for current models.
The trust machinery is the real lesson. The inbox log, the confidence threshold and the reply-to-correct loop are standard reliability patterns, and they matter more than the choice of tools. Most DIY AI automations skip logging, then get abandoned the first time something files oddly and nobody can tell why.
Memory, compute and interface as separate layers is the right architecture for anything you want to keep for years. It is the same reason to avoid storing your only copy of anything inside one chat app's built-in memory.
Small, scheduled outputs with hard word caps are a good counter to the usual failure of AI summaries, which is that they are long and unread.
What it leaves out
This is a task and relationship tracker, not a knowledge base. The four buckets cover people, projects, ideas and errands. There is nowhere for the articles, videos and documents you learn from, and no way to ask a question across what you have saved. Retrieval is left to Notion's own search and the digests.
The fix button is a fourth automation. He totals the build at three zaps, but catching a "fix:" reply in a thread, finding the original record and moving it between databases is its own workflow, and the hardest one to get right. Expect to spend extra time here.
Model confidence scores are soft. A self-reported number between zero and one is not a calibrated probability. The 0.6 threshold is a starting point you should tune by reading the inbox log for a couple of weeks, not a setting to trust blindly.
Privacy and cost get no airtime. Every thought passes through Slack, Zapier and a model provider. If the Slack workspace belongs to your employer, its admins may be able to see that channel, so use a personal workspace. Paid plan limits and per-task costs on Zapier also change over time, so check current pricing before you build around paths and schedules.
The pitch is partly a funnel. The prompts and worksheets are on his Substack, so the video gives the design but not the exact classification prompt.
How this applies to memory across AI tools
His memory-compute-interface split maps directly onto the AI tools you use. If memory lives in a store you control, any assistant can sit on top of it, and changing assistants does not mean starting over. That is the argument behind persistent memory across tools. Notion also offers an MCP server, so the databases from this build can be read by MCP clients, not just by Zapier.
The gap is the content side. Jones' loop captures your own thoughts. What you watch and read still needs a home. BrainTube covers that half: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base, exposed as an MCP server that Claude, ChatGPT, Cursor and other MCP clients can search. Running both keeps your actions in Notion and your source material searchable from the same assistant. There is also a side-by-side comparison with Notion.
Who should watch it
Watch it if you are not an engineer, already live in Slack and Notion, and keep losing follow-ups and half-formed ideas. The stack is at 5:31 to 7:32, the building blocks run from 9:00 to 17:00, and the principles from 17:00 to 24:30 are worth hearing even if you never build this exact system. The step-by-step setup is at 24:30 to 26:30. Skip it if you want a way to search and reuse what you read and watch, or if you prefer local-first tools: the opening five minutes are motivation, and the build is cloud-only by design.
Frequently asked
- Can I build an AI notes system without code?
- Yes. Nate B Jones' version uses Slack for capture, Zapier or Make for automation, Claude or ChatGPT to classify each message, and Notion databases for storage. He estimates about an hour to set up. Budget extra time for the correction workflow and for tuning the classification prompt.
- Why use Notion instead of Obsidian for AI automation?
- Cloud automation tools such as Zapier can write to Notion through its API directly. Obsidian stores local markdown files, so a cloud workflow needs a sync layer to reach them, which adds setup and failure points. Obsidian is still a good choice if you want local ownership and are comfortable with that plumbing.
- How do I stop an AI from misfiling my notes?
- Ask the model to return a confidence score with each classification and hold anything below a threshold for review instead of filing it. Log every decision with the original text so you can trace errors. Make corrections take one step, such as replying in the same thread.
Keep videos like this one searchable
BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.
Go deeper
More video notes
- Build an Obsidian BACKEND Not a Second Brain! · Eric Michaud
- Complete Guide to building an AI Second Brain with Claude · Ray Fu
- The NEW Way to Build a Second Brain w/ Claude (2027) · Nick Saraev
- Build a Self Improving Claude Knowledge Base · Paul J Lipsky
Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.