Raw dictation: AI as the new interface for knowledge work (2026-10-01)

Raw, unedited dictation from Rufus. To be tidied into a post/summary/rant for the team (see bead lip-jmu).

Source: Claude Code desktop session b2834ca1-d7ee-44ab-87e1-6fd4b033dc5f (started 2026-10-01, working dir ~/src/life-itself). Search for that ID to find the original thread.

I'm wanting to explain to my team about the revolution in work of AI at the more basic level, like not the obvious point also just of automation of tasks, but about, at least for knowledge work, the shift in interface. And so the basic point is that for a long, long time, when we're doing knowledge work, we built applications to make our knowledge work easier. So, I mean, just the obvious would be built like a word processing interface where you could type and you had what you see is what you get and you had Google Docs and you had spreadsheets and you know, almost every other product. Like let's think of Asana or GitHub or anything like for task management, companies would spend billions of dollars or hundreds, tens of millions of dollars building interfaces and products for you to basically manage your tasks more efficiently in your way.

And in a way, what I'm trying to say is that for a lot of work, the AI interface of what I'm in right now, let's say Claude, is your interface, your control plane for most of what you're doing. So what I'm getting at is, to give an example, and I was doing it a lot, which is I noticed the other day I'm with Astra and I said, oh, you know, Astra, what are your tasks? And what she's done is she's brilliantly used Claude, but she's used Claude to get the tasks and then put them in a GitHub issue. And then she's reading the GitHub issue. And I'm like, no, no, no, no, no. You want to be in Claude. It's a more efficient interface and it's bespoke. And it does two things that are amazing.

One is you can talk to it. And talking for most people is much more efficient than typing. Like you typed because you had to, because you couldn't dictate, because dictating was rubbish and slow and didn't work. But you actually want to talk. And the other advantage of talking, even if it's imprecise, you tend to be able to give a lot more context. You tend to actually express more than if you're typing, because actually typing is a bit painful, you don't type more. Now there's times when you make mistakes and you overdo it, but in general one of my tips, I've even seen with Sylvie today was like: no, talk, don't type, even if that means it's a bit messy. And even I noticed Astra and Sylvie both doing this, going in and editing what they'd written, like "oh no, I should correct that", and I'm like, no, no, don't do that, most of the time it's just not worth it unless it's something very important. Just dictate more.

And so what I'm getting at is a two-way interface. It's both an interface in, that you can speak to it and dictate to it. Whether you're, by the way, in ChatGPT, this works just as much. I'm using Claude because I believe the majority of people here are using Claude. But you can dictate to it, and it can create an interface back to you that's normally going to be better than anything that you can get from an application almost nowadays. I mean, in certain exceptional circumstances it won't, but most of the time it can. And why? Because your AI can build HTML. It can build any interface you want on the right-hand screen for you in real time. Do you want a graph? You want a graph view instead of markdown bullet points? I've got it. You want this or that? Now, there's a limit to that, but for a lot of knowledge work it can just build whatever you want or extract whatever you want in terms of information. It can output a table for you. It can output a graph, whatever you want.

And noticing how I see Sylvie doing things, even I then sat with her actually for the first time in quite a while to actually work together the last few days, and seeing what Astra was doing, I was thinking, oh, this is not obvious. And it wasn't obvious to me, but it's like, this is the way to go.

And what it also means is that things like GitHub or Asana or Google Docs really just become filing cabinets. In general, you're never going to go and look at those documents anymore, really. You might want to generate a presentation when you give a presentation. Though I even think the future of presentations is going to change a lot. But in a way, unless we need real-time collaborative writing or we're trying to really proof-edit something together, okay then we need a Google Doc, but a lot of the time we don't need those things. We humans are not going to go in and work in those places anymore. And this is also true of task management. You're not really going to do task management by going and reading a task issue and filling things in and ticking boxes, which is what we all learned to do. What we're going to do is dictate: "I completed this", "tell me what my next issues are", "analyze the dependency graph", "give me a plan of work for the next five hours", "give me a plan of work for the next two days". Things that were inconceivable to have bespokely done for me in an intelligent way, unless I was some very senior executive with two full-time assistants whose job was to understand me and understand what I was up to.

So that really changed the game. And the question you then almost ask of your tools like Asana or GitHub or whatever is: do they work with AI? You don't really care whether they're actually that good for you as a human. You care: are they efficient and reliable and well-structured for AI? And by the way, that's why I'm using Beads, for example, which you might not even know what that is. And the beauty is I've never even really used Beads from the command line. I don't really understand how it works, but I know it works for my AI extremely well.

Now, this doesn't mean that you don't need to understand anything about how your underlying tools work, because there are moments things go wrong and you want to know what tools are available. But in general, this is a real change. So, for example, I used to use GitHub all the time. I don't really go to GitHub anymore. In fact, I've almost also stopped using GitHub for tasks because it's kind of annoying with my AI. It's not very efficient. I explained to Astra: if you want to write into GitHub, it has to write to the GitHub API and it takes time and it has to do tool calls and it burns my tokens and it's slow. And it doesn't really have a well-structured dependency graph of tasks, which is what I want. By that I mean what tasks depend on what tasks, which depend on what tasks. In GitHub I'm forced into this structure of either an issue or a checklist item in an issue or a parent issue, but it's not very smooth between those layers.

So what I want to get at is that this is the game.

Part 2: How does the AI know where to look? Index, local disk, cloud (raw)

Added later the same day, also raw dictation.

If I'm using my AI as this kind of way to go and search for material, like issues or tasks to do, how does it know where to go look? That's why I said we have a central catalog or index that is listing all the initiatives that we're working on and where their repositories are, their written materials, and where their tasks are.

Now, I also want to emphasize, I don't normally always work centrally. I might work centrally to say, what's my work for this week? But normally what will then happen is I will say, oh, I need to go work on the design of the website. The design of the website takes place in the design repository, or the design folder in Drive. I'm going to go and build an AI session that's just working there. I don't need to load everything else. So I might open my AI session and say I'm working in the design folder, show me the tasks here, because then I don't need the tasks from everything else.

But if you are asking the question, Fae, at the beginning of my day or the beginning of the week, or I'm Uriel or I'm Fae and I want to see an overview of what's been going on or what's next, I do need that index. I then tell my AI, go read it and then go look in those places.

The last thing I'll say is if I'm working on my local disk, which has the efficiency that a call to my local disk is ten times to a hundred times to a thousand times faster than calling out to the internet. If you think of what GitHub repos are on my local disk, they're just like a cache of what's up on GitHub. They're the same thing as was on GitHub, but locally. I've checked out most of the repos on GitHub onto my local machine because I can then work locally more quickly. But I could be writing back and forth.

Having said all that, the main thing that doesn't work very well remotely is Beads. Beads works by running a command line on your local disk.

It's also true that when Claude runs in the cloud, what it's actually doing is cloning a repo. When you say in Claude Code to work on a repo in the cloud, it's actually git cloning that repo into a sandbox, a virtual machine up in the cloud, working there and then saving things back. And it's doing that because it can use all kinds of command line tools on an actual machine that it can't use when it's just talking to the GitHub API. It can search files very quickly, and it does not do that by the AI reading those files. It uses tools that have been built by humans over decades that are very good at searching files. That's why AI is so efficient: it has lots of tools it knows how to use. And that's why you might want to use it in the cloud.

What's cool in the cloud is that in Claude you can even select multiple repos that you're working on, and you can tell it to install Beads in the cloud. You can actually hack it and tell it, install Beads as a command line tool. It won't install it by default, but you can get it to install Beads and then work on Beads in the cloud. The one issue it will have, and I'm getting a bit technical here, is that it will have a certain problem doing what's called a Dolt sync, for reasons we can get technical about. But other than that, it will work fine, and you can just ignore that error most of the time.

Notes / key points (part 2)

Cross-initiative questions ("this week", "what's next") need a central index: initiatives, their repos/written materials, and where their tasks live. The AI reads the index, then goes to those places.
Focused work doesn't need the index: open a session in the relevant repo/folder (e.g. design) and ask for tasks there only.
Local checkouts of GitHub repos are a fast cache (10x to 1000x faster than going over the network); sync back and forth.
Beads is local-CLI-shaped, so it works poorly remotely.
Claude in the cloud = git clone into a VM sandbox, so it can use real command-line tools (fast file search, grep etc.) rather than the GitHub API. The AI's efficiency comes from the tools it uses, not from reading every file.
In the cloud you can select multiple repos and tell Claude to install Beads as a CLI tool. Known issue: Dolt sync errors; mostly ignorable.

Part 3: Where do tasks live, and how does a colleague (Fae) set Claude up? (raw)

Added later the same day, also raw dictation.

Planning repo, okay. I have my own tasks on onboarding, how do they link up? I think you asked the question: in general most tasks, whether created in GitHub or in Beads in a repo, are at the level of the lower-level initiative. They're not at the planning level. In the case of the Seeds of Renaissance rebranding we didn't really have a clear initiative where that went. At one point I was doing it in the comms repo, and this is something we need to clean up a bit. I agree there's a bit of messiness about where these live. But in general, we've got to pick something a bit concrete you're working on.

Let's say you were looking for tasks in ops, Fae. You would say, I'm going to go look in the GitHub ops issue tracker, or in the Beads. At the moment we're running both, we're in an in-between state where we haven't finalized. Go and look through those two and find me things I'm assigned to.

Now you had another question about how I set up Claude so that I have access. This depends whether you're running in the cloud or locally. If you're running locally, you can still tell it to go off to GitHub and it can go online. But if you're looking for it to look through files on your local disk, you need to have cloned those repos. So if you're trying to work locally against local files, then you need to have cloned them locally. But you can run locally and talk online, as I could demonstrate. That's a little bit subtle and probably Claude can give a better answer here.

Ask to Claude: write this raw outflow again as another section; play back to me what the question is I'm being asked and how I'm trying to answer it; and, if you were going to generate a short tutorial or guide, or link to any guides, explain to Fae (my colleague) how she would work across multiple repos, and particularly access Beads or files, either online or cloned locally, from Claude Code, running in the Claude desktop app, not on the command line.

Notes / key points (part 3)

Tasks mostly live at the level of the lower-level initiative repo (ops, comms, design…), not in planning. The planning repo is the index/overview.
Messiness to clean up: rebrand tasks have been split across comms and planning, with no clear home initiative.
Currently two trackers in parallel (GitHub issues and Beads), so "find my tasks in ops" means checking both.
Local session + online GitHub works; local session over local files requires the repos to be cloned.

Notes / key points extracted (for the tidy-up)

Old model: billions spent building apps/interfaces (Word, Docs, Sheets, Asana, GitHub) so humans could do knowledge work.
New model: the AI chat (Claude/ChatGPT) is the control plane. Apps become filing cabinets / storage.
Two-way interface: dictate in (more context, faster than typing; don't edit, just talk more), bespoke UI out (HTML/graphs/tables rendered in real time).
Example: Astra using Claude to fetch tasks, writing to GitHub issue, then reading the issue. Should stay in Claude.
Task management becomes conversational: "I did X", "what's next", "analyze dependency graph", "plan my next 5 hours / 2 days". Executive-assistant-level service for everyone.
New tool criterion: is it efficient, reliable, well-structured for AI? (Beads: dependency graph, local, fast, cheap in tokens vs GitHub API tool calls.)
Caveat: still understand your underlying tools for when things break.
Still need Google Docs for real-time collaborative writing/proof-editing; presentations will change too.