Three years ago i wrote about strategies for coping in the information age. The core problem was volume: too many newsletters, podcasts, tweets, and articles competing for limited attention. My solution was a pipeline — Gmail filters to triage inputs, Readwise to capture highlights, Snipd to unlock podcasts, and Notion as the central database pulling it all together.
That system worked well. But the world it was built for has moved on.
In 2023, the bottleneck was ingestion. You had access to brilliant people and their ideas but you couldn’t capture and process it all. The tools i recommended were essentially elaborate filing systems — ways to capture and categorise information so your brain could do the hard work of making sense of it.
In 2026, that hard work — reading unstructured information, applying what you know, producing something useful, checking it’s right — is increasingly something AI can do alongside you. Doug O’Laughlin at SemiAnalysis describes this as the READ-THINK-WRITE-VERIFY loop: the basic cycle that defines almost all information work. Whether you’re writing code, analysing financial statements, or drafting a board paper, you’re running some version of this loop hundreds of times a day.
The important thing to understand is that – as Andrej Karpathy puts it: “The main effect is that I do a lot more than I was going to do because I can code up all kinds of things that just wouldn’t have been worth doing before.” Byrne Hobart makes the same observation: AI “has reduced the amount of time needed for researching any given piece… but has also expanded the scope of things worth researching.” The frontier of what’s possible has moved, not just the speed at which you reach it.
The question has shifted. It’s no longer “how do i capture everything?” It’s “how do i get AI to help me process it?”
This article is an update. Same practical lens. But the toolkit has changed fundamentally — and so has the way information flows through a personal knowledge system.
Why i left Notion for Obsidian
In 2023 i called Notion “mission critical” and predicted it would be one of the most valuable public companies in the world when it IPO’d. I still think Notion is a great product. But in January 2026, i migrated my entire knowledge base — roughly 10,000 files — out of it.
This was driven by the acceleration of AI over the Christmas 2025 break where Agentic Command Language Interface (CLI) tools, have become genuinely – and slightly mind-blowingly – useful. Ben Horowitz captured the moment well: “Over the winter break, it turned a corner where really good programmers were going, ‘Whoa, oh god, this helps me.’ I can’t remember any kind of technology where all of a sudden you wake up and the whole world just changed like that.” And these AI agents work best with markdown files.
Notion stores your data in a proprietary format behind an API. To get anything in or out programmatically is a pain. Obsidian, by contrast, is just a folder of markdown files on your hard drive. No database. No API. No lock-in. Your notes are plain text files that any tool — including an AI agent — can read, search, edit, and create directly.
This matters because the most capable AI tool i use — Claude Code — is a terminal-based agent that works with your local file system. It reads files. It writes files. It runs scripts. If your knowledge base is a folder of markdown, Claude Code operates on it natively. If it’s locked inside a proprietary database, it can’t.
Steph Ango, the CEO of Obsidian, calls this “file over app” — if you want digital artefacts that last, they must be files you control, in formats that are easy to retrieve. I’d add: if you want AI – and specifically CLI agentic tools – to work with your knowledge, it needs to be in a format AI can actually touch. As Bill Gurley put it on a recent All-In episode: “It’s not open source or closed source anymore. It’s open data or closed data.” That’s the divide that matters now.
The migration was scary (notion was my “second brain” after all). Over a weekend i had Claude Code import and fundamentally restructure all 10,000 files — normalising metadata (adding YAML tags to everything), assigning categories, removing duplicates. I ran multiple agents in parallel while i was on a hike. Came back to find the work done. Cost: over £200 in AI tokens. But it was worth it to me.
Claude Code: the bit that changes everything
So now Obsidian is where things live, while Claude Code is the thing that does stuff with them. Steph Ango captured the convergence in a tweet: “English is now a programming language. You can write programs just by writing plainly in Obsidian, and run your program using Claude Code.” And that’s not a BS metaphor – It’s literally what i do most days from c. 5am.
For those living under a rock, Claude Code is a terminal-based AI agent from Anthropic. It’s not a chatbot. It has full access to your computer — reads files, writes files, runs scripts, browses the web. You tell it what you want in plain English and it works out how to do it. Think of it less as a tool and more as a very capable intern who happens to live inside your terminal and never sleeps (no offence intended to interns).
I use it for almost everything. In my first month i logged over 2,500 sessions and 16,000 messages. It touched 13,000 files. I’m not unusual — SemiAnalysis estimates 4% of all GitHub commits are now authored by Claude Code, heading toward 20%+ by end of 2026. But the raw numbers aren’t the interesting part. The interesting part is skills.
Skills are reusable instruction files — basically markdown documents that teach Claude Code how to do specific jobs. I’ve built about 30 of them. When i say “turn this PDF into a PowerPoint,” it loads a skill that knows “my” formatting rules, layout logic, colour palettes. When i say “publish this essay to my website,” it loads a skill that knows my WordPress API credentials, taxonomy structure, and how to generate SEO metadata. The podcast transcript you might have seen on my site? Claude Code fetched it from YouTube, formatted it as an Obsidian note, and filed it — all from one instruction.
What makes this different from ChatGPT or any other AI assistant is that it actually does things. It doesn’t give you a helpful suggestion and wish you well. It ingests a podcast transcript, normalises the metadata, files it in the right folder, links it to related notes, and moves on to the next job. Tasks that used to take five tools and thirty minutes now happen in a single conversation.
The compounding effect is where it gets slightly absurd. Skills improve over time — there’s a reflect skill that analyses each session and proposes improvements to its own instructions. Hooks fire after every edit to validate metadata. Scripts run hourly to process incoming content without any human involvement. Between January and February 2026, Claude Code added over a million lines across my projects. The best analogy i can think of: you’re not using a tool anymore. You’re managing a small team that happens to be software.
How information gets in
The front end of the funnel has changed a lot since 2023. The principle is the same — be deliberate about what you consume, capture the best bits — but the tools are different and the automation is much deeper.
Obsidian Web Clipper has replaced most of what Readwise used to do for articles. It’s a browser extension that clips any web page directly into your Obsidian vault as a markdown file, with metadata pre-populated. I still use Readwise for Kindle highlights and some RSS feeds, but for day-to-day article capture, the Web Clipper is faster and cuts out the middleman.
Voice memos are now a proper pipeline rather than something i forget about. I record thoughts on my iPhone — walking, in a cab, wherever — and iCloud syncs the audio to my Mac. A background script picks it up, sends it to Google’s Gemini API for transcription, extracts a title and topics, creates an Obsidian note, and links it to that day’s daily note. The whole thing is hands-off. I record, i forget about it, and the thought shows up in my vault later that day. For longer-form dictation — meetings, interviews, extended thinking — i use Otter AI, which does real-time transcription with speaker identification.
Podcast transcripts used to be a pain. In 2023 i relied on Snipd, which was decent for clipping segments but couldn’t give you the full transcript. Now i just ask Claude Code to fetch the transcript from YouTube (most podcasts have a YouTube version), format it as an Obsidian note with highlights and key quotes, and file it. Takes about 30 seconds.
OpenClaw handles the email layer. This is an open-source AI assistant i run on a remote server, connected to my Gmail. Every morning it sends me a briefing via WhatsApp: calendar for the day, important emails from the last 12 hours, and a digest of which newsletters are actually worth reading. It auto-labels incoming mail, archives noise, and flags things that need a response. Every evening it sends a recap. The newsletters i don’t read get archived automatically. The ones OpenClaw flags as relevant to what i’m working on get surfaced. This has replaced all the Gmail filter rules from my 2023 setup — and it’s considerably more intelligent, because it understands context rather than matching keywords.
What happens to the information
In 2023, the “processing” step was mostly me. I’d sit down with my Notion database, read through highlights, mind-map connections in Xmind, and manually synthesise things into drafts. The tools helped with capture but the thinking was entirely manual.
Now, a lot of the mechanical processing happens without me.
Automated categorisation runs hourly via background scripts. When Readwise syncs a new article or tweet into my vault, a script analyses the content, assigns up to five topic tags, and removes the “inbox” label. Books get normalised with genre and description pulled from the Google Books API. Podcasts get their show name, guest, and episode number extracted from the title. By the time i look at my vault, most new content has already been filed and tagged.
Claude Code’s parallel agents handle the heavier processing. When i had 436 unprocessed tweets after migrating from Notion, i didn’t sit there categorising them one by one. I launched six parallel agents and they worked through the lot simultaneously. The same pattern applies to any batch operation — normalising hundreds of book entries, cross-referencing meeting notes against project files, auditing vault health across 4,000+ notes.
NotebookLM from Google has become my go-to for synthesis. I load a set of source documents — meeting notes, research papers, articles on a topic — and it generates an audio summary in podcast format that i can listen to while doing something else. More recently, it produces visual summaries too — diagrams and relationship maps that i can screenshot and bring into Claude Code to turn into proper presentation graphics or Figma designs.
Nano Banana (Google’s Gemini image generation) fills the gap for quick visual content — infographics, diagrams, conceptual graphics. Combined with NotebookLM’s visual outputs, i can go from raw research to a polished visual in minutes rather than hours in Figma.
The pattern across all of this: the mechanical work — tagging, filing, formatting, initial synthesis — is increasingly automated. What’s left for me is the judgment work: deciding what matters, what connects to what, and what’s worth writing about. As Karpathy puts it: “Programming feels more fun because the fill-in-the-blanks drudgery is removed and what remains is the creative part.” The same applies to knowledge work more broadly. Which, frankly, is the bit i actually enjoy. Steph Ango makes a similar point about note-taking specifically: “People have asked me if this could be automated with language models but I do not care to do so. I enjoy this process. Doing this maintenance helps me understand my own patterns.” He’s right. The goal isn’t to automate thinking. It’s to automate everything that isn’t thinking.
From notes to published work
The output side has changed as much as the input side. In 2023, Notion doubled as my drafting tool and i’d manually copy content into WordPress, Substack, or Twitter. Now, Claude Code handles most of the last mile.
Writing still starts with me. I draft in Obsidian — usually in the vault root as a plain markdown file. But from there, Claude Code takes over. It can generate SEO metadata (title, description, focus keyphrase), push the article to WordPress via the REST API, and handle the formatting. For my website, it knows the theme, the taxonomy, the font stack. I tell it to publish and it publishes.
Presentations are the most dramatic change. In 2023 i used Figma for visuals and manually built slides. Now i can hand Claude Code a plain-text article or a set of notes and it produces a branded PowerPoint — correct layouts, colour palette, formatted charts. It has a skill for this that understands OOXML, the underlying format PowerPoint uses. The same workflow works for Word documents and PDFs. Markdown in, polished document out.
Visual content comes from a combination of Midjourney (which i still use for cover images and portraits), Nano Banana (for diagrams and infographics), and NotebookLM (for visual summaries that get refined in Figma or turned into presentation graphics via Claude Code). The cover image for this article, for instance, was generated by Midjourney — same as in 2023. Some things don’t change.
Publication channels are mostly the same: WordPress for long-form essays, Substack for newsletters, LinkedIn for professional reach. The difference is that getting from “finished draft” to “live on all platforms” went from an afternoon’s work to about five minutes.
The updated process
The pipeline i described in 2023 — Information & Inspiration → Capture & Aggregation → Analysis & Visualisation → Drafting & Editing → Publication — still exists. The stages haven’t disappeared. What’s changed is that two orchestration layers now span across what used to be entirely manual steps.
Claude Code stretches from Capture all the way through to Publication. It fetches podcast transcripts (capture), runs parallel categorisation agents (processing), synthesises notes into drafts and presentations (output), pushes content to WordPress (publication), and validates metadata via hooks (verification). It’s the connective tissue.
OpenClaw spans Information through to Processing. It triages my email, surfaces the newsletters worth reading, sends morning and evening briefings, and runs automated jobs and a lot of other stuff (I’m trying to get it to check me into flights and pick seats etc). Where Claude Code is a tool i direct, OpenClaw is mostly a tool that runs while i sleep.
Underneath both sits Obsidian — as the persistent database layer for all my thoughts, ideas, inspiration and learnings, that everything reads from and writes to.
The READ-THINK-WRITE-VERIFY loop
Stepping back, the system i’ve described maps onto a simple model that applies to almost all information work:
- READ: Ingest unstructured information. Web Clipper, Readwise, voice memos, podcast transcripts, OpenClaw email digests — all of these are READ tools. They take messy, scattered inputs and get them into a structured format.
- THINK: Apply domain knowledge. This is where NotebookLM synthesis, Claude Code’s parallel analysis, and — critically — your own brain come in. The automated categorisation and cross-referencing is THINK at the mechanical level. The judgment calls are THINK at the human level.
- WRITE: Produce structured output. Drafting in Obsidian, generating presentations, publishing to WordPress, creating visual content — all WRITE.
- VERIFY: Check against standards. Claude Code’s hooks validate metadata on every edit. The
reflectskill audits its own performance. Vault health scripts check 4,000+ notes for consistency. In a professional context, this is where peer review and quality assurance live. And this step is non-negotiable — Andrej Karpathy describes LLMs as having “amusingly jagged performance characteristics — they are at the same time a genius polymath and a confused grade schooler.” You cannot skip verification just because the agent sounds confident.
What’s changed between 2023 and 2026 is which parts of this loop require a human. In 2023, you were doing all four stages manually, with tools helping at the margins. In 2026, READ and WRITE are largely automated. VERIFY is increasingly automated. THINK — the application of judgment, context, and taste — remains the human part. And honestly, that’s as it should be.
The way i think about it: there’s an outer loop and an inner loop. I run the outer — what to pay attention to, what matters, what to create. AI runs the inner — capture, categorise, format, publish, validate. The inner loop gets faster every month. The outer loop is where the value lives.
Final thoughts
Three years ago i ended with some predictions. I said Notion would IPO as one of the most valuable companies in the world (it hasn’t, and i’ve since left the platform). I said AI would have profound impacts on the future of work (this one landed, rather emphatically). I said large businesses would need to get more agile with IT procurement (they’re trying, slowly).
Here’s my updated prediction: within two years, the system i’ve described — a personal knowledge base of plain files, orchestrated by AI agents, with automated capture and processing pipelines — will be table stakes for anyone doing serious information work. Platforms that embrace open data will be winners. Close data players will be losers.
Most corporations are way behind the adoption curve on these tools – and in part for good reason (information security) – but I’d recommend playing with the tools in your personal capacity. It’s cheaper than you’d think (obsidian is free!), very fun and rewarding, and it will help you better prepare for and visualise the future of (knowledge) work.
A caveat, though. Simon Willison — one of the sharpest observers of AI tooling — has flagged prompt injection as the risk “most likely to result in a Challenger disaster.” These systems are powerful but also easily manipulated. The more autonomy you give agents, the more important it is to understand what they’re doing and where the guardrails are.