One of the problems I kept running into was Claude’s memory, I’d settle something on the Code surface at my terminal, but the Chat surface had no idea what had been settled so i had to reexplain everything that had been settled. I also had a growing pile of projects I wanted to run in parallel. There was a lot of talk online where people were following Karpathy’s idea of a “second brain” built in Obsidian. I was already comfortable in Notion, and the MCP connector already existed, so I tried that route instead. It slowly turned into a simple loop that’s kept me in control of a lot of projects at once, on the move and at the desk. Notion also has some benefits with working in teams, using alternative models (if Claude’s not your cup of tea), the benefits are explained below but if your chatbot has an MCP tool you can use this as your storage system.

The loop I’m sharing is simple: /start → work → /close. Each cycle is a session, and I’m now at around 500 of them across my workspace. It’s gone through a lot of iterations, and I’ve decided to share it, along with some of the other projects it’s helped me build. None of it is Claude-specific either, that’s just the one i use most, if your chatbot has an MCP tool, Notion works as the storage layer.

Basically every interaction with Claude i open with /start, it gathers context from Notion including a list of to-dos. A reminder is posted on what I was working on last session, which tasks are still open, and what I said I’d do next. I don’t need to write any of it in because Claude write’s it in at the end of a session. At the end of the session i type /close, and it writes the summary, ticks off what got done, and sets up tomorrow.

This first iteration is built so you don’t need any code, it’ll work right from the chat surface (or cowork) from your phone or desktop. You’ll need a few things: Claude Subscription and a Notion Subscription (about £25 in total). Below you’ll find some expose on the “surfaces”, why notion not obsidian, my top tips and more speel on the benefits of what’s referenced as a “second brain” but if you just want to get going the template is here:

Link the connector in Claude to Notion, start up a new project and paste in the provided prompt. Use Opus or Sonnet ideally, i’ve tried it with Haiku and said “install this” and pasted the wrong URL from the Notion page and it got roughly the right thing. The way the template is set up is based on the loop itself, you’ll hopefully go through around 3 sessions to get your project off the ground and there’s extension tasks if you want to set up the code surface and cowork surface too. I’ve been banging on about “surfaces” these are just my colloquial name for the points at which you can interact with Claude.

The surfaces — Chat, Cowork, Code

Claude runs in a few places, and they are not alwats interchangeable. Most of the confusion people hit when they wire Claude into their tools comes from expecting the surfaces to behave like one another. Each have their own capabilities but all of them can reach notion. At the time of writing this is roughly the shape of the surfaces but The rough break down is below:

Chat
The first time you come across any LLM it’s likely through this kind of surface, claude.ai on the web and your phone. No file system, no terminal — but it has the same connectors and the same skills as the desktop app, so once you’ve connected Notion, Chat can read and write your workspace directly.

Cowork
This and chat are slightly different but there has been a pus is the Claude desktop app. Everything Chat has, plus a sandboxed shell and the ability to read and write files in a folder you point it at — so it produces real deliverables. Same loop, more reach.

Claude Code
Runs in your terminal and IDE. The deep end: local files under git, skill commands as real files, hooks that fire at session end, Python, a shell. If you want Claude to push a paper to your reMarkable or run a migration, this is the surface. You set it up later, when you want it — as a single pasted prompt.

As Anthropic pushes towards more agentic work Cowork is looking like the leader and Chat may be retired whilst you’re reading this but at the time of writing they are still seperate entities so deserve their own section. As I’ve said, the connectors are shared regardless of surface.

What the Notion connection actually gives you

Installing the Notion connector gives Claude access to the Notion MCP, this just allows Claude to access your Notion, read from it and write too it. The reason we need this is that Notion is seperate from Claude and so we can set it up to be both persistent and well structured.

Persistent in the sense that what Claude writes to Notion outlives the conversation it was written in. Write a session summary on Monday from the terminal, and Claude can read it back from your phone on Wednesday’s commute. The information is written at /close and is exactly what makes the /start moment possible.

Structure means Notion’s own shape — pages inside databases inside hub pages — this maps cleanly onto how i like to run a project. The structure I settled on was “a current state” backed up by a list of open tasks, and a session log, this is the general shape of the project structure, it allows Claude simple navigation rather than a single long file it has to re-read from the top.

Underneath both of those, the mechanism is just Claude being able to write to a page or update it, and read it back later, so it can keep a consistent understanding of what you’ve already talked about. The read and the write are the tools; persistence and structure are what you will hopefully feel. The advantages of notion are explained below but really i used it because it was already a part of my workspace.

Why Notion and not Obsidian

The honest answer first: Obsidian is excellent. Local, fast, Git-versioned, free, with a plugin for almost anything. If your priority is owning plain-Markdown files with no subscription, Obsidian is the right call for you. But for an AI that needs to read, write, and reason across a multi-project, multi-surface workspace, Notion has some structural advantages over Obsidian.

• It’s always online, from anywhere. Claude on your desktop, a scheduled routine at 3am, Claude on your phone — all reach the same Notion workspace at once, no mounted vault, no syncing to check. If Obsidian’s local-first model is what makes it good for you then that’s cool.
• Databases are first-class. Typed properties, filters, relations, multiple views. Claude writes a row and it’s immediately queryable by priority, project, or date. Dataview parses frontmatter, but that’s text interpretation, not structured storage.
• Cross-database relations. One Global To-Do database, linked into every project. Claude writes to one canonical source; it appears everywhere that references it.
• Shareable templates. A Notion page duplicates worldwide in one click. The whole install model depends on it: you duplicate a page, tell Claude the URL, and it sets itself up.
• Block-level writes. Append to one section, insert at the top, update a status — without touching the rest. Flat Markdown is append-or-overwrite.

None of this makes Obsidian wrong. If offline-first and plain files matter more to you than structured, always-online data, Obsidian wins. But this workflow is built around agents that run on a schedule and templates that install themselves — and for that, Notion is the better foundation.

What becomes possible

Once Notion talks to Claude, you’ve got a working brain — and it’s a specific, small thing, not a vague promise. Every project gets its own hub page holding: a Current State page, the one paragraph Claude rewrites at the end of every session so the next one opens already knowing where things stand, instead of re-reading everything that came before. An inline Session Log — one row per session, what got done, what’s next — the thing /start reads from and /close writes to, it’s a version of Git, you can probably iterate on this and make it far better than my bare bones DB. From the “hub page” i included two filtered views onto databases that otherwise live at the root: a Global To-Do DB, cut down to just this project’s open rows, and a Notes DB, filtered the same way. That’s how i shaped the project, you might have something different but it should run from Chat on day one without any code so if you want to try it out then it can be found at my marketplace. Because Claude reads all of it directly, you never copy-paste context into the chat — ask what’s open on a project and it queries the filtered view itself.

After this i’ll start to post the rest of the repertoire that i’ve built out, this include supabase integrations as well as what I call meta agents: not an autonomous background process, but an ordinary interactive chat that takes on a persona and a toolset, and orchestrates work across several projects underneath the conversation. That’s still forward thinking, a little teaser, 101 Factory, is free and due out next week. Add a second surface today, though, and it’s the same loop, re-pointed: Claude Code as a single pasted prompt that wires up files, git, and hooks itself.

Setting it up

The first init is a conversation rather than a config file or a .exe. Duplicate the Notion bundle into your workspace, connect Notion to Claude, and paste one setup prompt into Chat or Cowork. (There’s a secondary entry through Code and Git if you’re that way inclined). Claude walks the Notion workspace, writes the grounding files, and mints the loop skills — you just confirm as it goes. It should take about ten minutes, no terminal. The install is set up so that you have to go through the loop a few times before it’s all fully installed, i estimate about 3 sessions, it can be done in one session but getting into the habit of opening and closing sessions is kind of the whole point; after that there are optional extension tasks if you also want the Code surface or Cowork wired in. Running costs are not negligible but fixed: a Claude subscription plus a Notion subscription, about £25 a month between them.

Use Sonnet or above for the smoothest first run. Before I trusted this enough to share, I ran it through the model tree end to end — not just reading the tutorial, but actually following it. The first Haiku attempt was rough: told to install it, it made a dozen tool calls, wrote very little, and told me it was done anyway. Persisted through that and it produced two projects – a Wild Sourdough Archive and a Pigeon Post. After the bug fixes in the prompts I ran it through, Sonnet, Opus and Fable which ran the whole thing cleanly, each picking its own two example projects to prove the loop end to end. Sonnet went with Bell Chords and a paper hygrometer; Opus adopted a neglected allotment, then took on a house extension; Fable restored grandad’s old Raleigh and then started an allotment of its own. Haiku, rerun after the fixes, got through it with a home-office redesign and a React component library the second. – i am considering running a project in a project and seeing what a Haiku’s React component library looks like.

One honest caveat: as of writing, this install path has been run by agents across every Claude model class, not yet by a stranger starting cold on their own account. The one real gap that turned up: it’s possible for a setup conversation to look finished while a skill silently failed underneath it. That’s exactly why there’s now a check built into the install itself, so you find out immediately rather than three sessions later.

The one thing to know going in: MCP connectors and custom skills on claude.ai need a paid Claude plan. Beyond that, there’s no separate per-page sharing step to remember — the connector is OAuth, and once it’s connected it sees whatever your account can already see in Notion. If a page comes back as invisible, the fix is to reconnect the connector or paste in your desired URL.

→ Notion Marketplace — Simple Claude Loop

The Code-surface starter: github.com/iniphi/notion-claude-loop (private while this post is still a draft — public the same day this goes live)

This is the entry point. Once it’s done, the rest of the lineup rides on the same brain:

• reMarkable × Zotero × Claude — your reading pipeline
• Zotero × Claude — your reference library, searchable by AI
• [more coming]