Laura Rosenberger

Category: Field Notes

  • How I set up folders so a new chat knows what it needs to know

    People get very attached to their chat threads. I see it constantly. Someone has one conversation running for a project, or a client, or a whole business, and it has been going for months. The thread is doing the job of a folder.

    I understand why. Everything that got you to a decision is in there, so deleting it feels like deleting the reasoning.

    It is also the worst place to keep it. Every time you send a message, the model reads the whole thread back. It sees decisions you made and then reversed. It burns through your usage allowance re-reading a conversation you had already finished with. And none of it is portable, because it lives inside one tool, in one account.

    The Solution

    I start with the folder. That is where the work lives, and the AI gets invited into it.

    The test I use is whether a brand new chat, with no memory of anything, could open that folder and be up to speed in a minute. If it could, the folder is doing its job. If it could not, whatever it is missing needs writing down.

    How I Use It

    Three habits do most of it.

    One folder per unit of work. A client, a project, a part of the business. Keep it at that level, because tasks come and go and the folder outlives them.

    Only the current version of anything lives in it. When there is a new version, I create a subfolder called 'archive ignore' and move the old ones in there. The AI reads the folder name and skips it, and I can still go and find the old file when I want it. This one is worth doing today. Most of the times I have seen AI working confidently from a stale document, it was because several versions of that document were sitting in the same place.

    And a plain note of what was decided and why. It only needs the handful of things that would otherwise live in someone's head, or in a chat thread from March.

    I've written before about getting each session to write down what the next one needs. That is the same habit, automated.

    Why This Has Impact

    The chat thread stops being precious. I open a new one whenever I want, because everything it needs is on disk and it reads it in seconds.

    The bigger shift is that the AI only ever knows what it can read. So the time I spend putting things into files decides how much of the work I can hand over. It has stopped feeling like admin knowing that this is going to make or break how well my AI works.

  • How to fix AI amnesia by connecting Claude to your to-do list

    It's becoming an increasingly common problem, a sort of AI amnesia: people produce a piece of work with AI, and then forget they produced it.

    I hear versions of it constantly. Someone spends ten minutes getting Claude to pull together the research for a meeting, then walks into that meeting a fortnight later having completely forgotten the document exists. Or they redo it from scratch, because asking again is quicker than remembering where the first one went.

    I think the effort used to be what made it memorable. An hour of real work leaves a mark in a way that two minutes of prompting doesn't.

    The Solution

    I worked this out with one of my coaching clients, who had exactly this problem. They were producing genuinely useful things in Claude and then losing track of them.

    Their to-do list app was already connected to Claude through an MCP connector, which is what lets Claude add tasks to it directly. So we used that: at the end of any piece of work, ask Claude to drop a to-do into the task inbox for the date the work is actually needed, with the file path to the document inside the task.

    The prompt is nothing clever. Something like: 'Add a to-do for next Friday to review these questions, and put the file path to the document you just created in the task.'

    How to Use It

    Two small details do most of the work.

    The first is the date. Put in before the day you expect to pick the work up. If Friday arrives and you're not ready, push it to the following Wednesday. The to-do is a reminder that the work exists, so it can move as often as it needs to.

    The second is asking for the file path rather than a link. Left to its own devices Claude will often reach for a sharing link, which is fine for sending to someone else but useless here. The file path is the thing Claude can open again later.

    Then on the day, you don't go and find the document. You tell Claude 'we've got a to-do today to review those questions, can you pull them up?' and it opens what it wrote. You never navigate to the file yourself, which means you don't have to remember where it lives or what you called it.

    Why This Has Impact

    The obvious benefit is that work turns up when it's useful instead of disappearing into a folder. Research done three weeks ago for a call is in front of you on the morning of the call.

    The less obvious one is what it says about where the constraint has moved. The hard part now is keeping track of what you've already got. I've started treating every AI output as something that needs a return journey booked for it.

  • AI note-takers feel fine on a video call and rude in person

    I have a note-taker running on almost every video call I do. It's completely normal now. Nobody comments on it.

    Put a phone down on the table between me and someone I'm sitting opposite, and it feels rude. I wouldn't do it. I've been trying to work out why, given it's obviously the same thing.

    Part of it is that people talk differently when something is recording. On a video call that's already priced in, because everyone is half-performing to a camera anyway. In person you get candour, and the offhand remark someone makes while they're thinking, and that's most of the reason to be in the room at all.

    The rest of it is just that it isn't a normal thing to do yet. There's no established etiquette, so it reads as odd, and being the person who does the odd thing changes the conversation before anyone has said anything. That part will probably fade.

    I think the best way round it is to not record the meeting at all. At the end, say something like 'let's just whip through what we talked about', and record that. Two minutes of the two of you agreeing what was decided and who's doing what.

    It loses the texture of the conversation and keeps the actions and the decisions, which is the part worth going back to. By then the real conversation has already happened, so the recording isn't changing how anyone talks. It's an easier thing to ask for too, because summing up at the end is something two people do together anyway.

    The phone still comes out. It just comes out once the conversation everyone came for is finished.

  • How I optimise my Claude usage limits by running tasks overnight

    Claude's usage limits come in five-hour windows. I kept bumping into them during the day, then sleeping straight through several windows every night. Capacity I'd paid for, completely unused.

    At the same time, I end most days with tasks that Claude could do without me. Build this tool. Research that topic. Tidy up those notes. None of them need me at the keyboard. They just need a decent brief.

    The Solution

    I tag a task 'overnight' in my to-do list app, and my Mac does it while I sleep.

    In the evening, a small script picks up anything I've tagged, turns the note into a job, and works out which project folder it belongs to. At 2am the Mac wakes itself up and Claude Code works through the queue – a fresh session for each task, in the right folder. By the time I sit down in the morning, the results are waiting for me.

    How I Use It

    The to-do note is the brief, so I write it as if I'm handing the task to a person: what I want, where the relevant files live, what good looks like. A lazy note gets a lazy result.

    A recent example: I came across this list of signs of AI writing (https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) and wanted to create a 'slop checker' skill to run over my writing before I publish anything. I wrote that up in a to-do, tagged it 'overnight', and went to bed. By morning Claude had built the skill and tested it against sample documents.

    Why This Has Impact

    The most obvious change: I run out of my Claude allowance far less during the day, because a chunk of my work now happens in windows I used to sleep through.

    The change I didn't expect: it's making me more specific. There's nobody at the keyboard at 2am to nudge things along, so the brief has to be clear enough that Claude gets it right without me intervening or guiding. That discipline is making me better at writing specific prompts, and at giving Claude access to the context that lets it run more autonomously than if I was babysitting it.

  • Building websites for agents, not humans

    I've built two websites recently. And I fell into a bit of a trap – I built them with humans in mind, not agents. How old fashioned of me!

    My target customer is consumer brands who need help with AI – getting up to speed with it, understanding what's possible, integrating it into their workflows. This isn't something that people in those teams would have previous experience with.

    For the person tasked with figuring this out, where are they going to look for answers?

    Most likely via ChatGPT or Claude.

    So I should have been building my websites with Claude in mind (rather than that person in my target company), to maximise my chances that my content will be surfaced in those chats.

    I've been following the guidance in this brilliant post by Tom Osman around creating llms.txt files, structured data, and comprehensive meta data.

    And asking Claude Code to integrate it for me, naturally.

    The web is going to serve two distinct purposes: a web for humans and a web for agents. Trying to make one website serve both would be a shame, so I hope the web will split in two. The human web can double down on beautiful experiences, thoughtful interfaces, content that's genuinely worth reading and stumbling upon. The agent web can be plain text, structured answers, Markdown, and can be where the mass-generated SEO blogs we've all learnt to scroll past can live. One is where you browse and discover, the other is where you pull down what you need.

  • AI Memory Across Sessions

    One of the reasons OpenClaw has captured everyone's imagination is the memory. It remembers you. Every other AI conversation starts from zero – close the window, lose everything, re-explain yourself next time.

    I've been working a lot in Claude Code, and have built myself a little skill to recreate this.

    How It Works

    I built a /handoff command that runs a structured wrap-up when I'm done working. It reviews the session and saves what matters to a set of persistent markdown files — a daily note, active threads, behavioural patterns (which it can then pull me up on…), and self-improvement notes.

    These notes are accessible for future sessions. Claude knows what I was working on, what's due, and what patterns to watch for.

    The whole thing is just a text file telling Claude what to do and what to save – but it is making using Claude Code as a general assistant for non-technical work just so much better.

    Why It's Useful

    It turns each session from a one-off into something that compounds. Less time re-explaining, more time working. And it catches things I wouldn't note myself — like when the same blocker keeps showing up or when I'm quietly avoiding something (ahem).

    It's not as seamless as OpenClaw's built-in memory. But I control what gets saved, where it lives, what gets forgotten and what my 'assistant' is able to do with this information.

  • AI Champions Are Being Set Up to Fail

    There's a new role quietly appearing across businesses: the AI champion. It's rarely a formal hire. It's usually someone already in the company (a generalist, a special projects person, someone who's enthusiastic and open-minded) who gets handed the AI brief on top of everything else they're doing.

    It's a crucial role. Arguably one of the most important in any business right now. And for the individual, it's an incredible opportunity to become the person who actually figured this out.

    But the support is basically zero.

    There's no playbook. No established methodology. No 'project management 101' equivalent for AI transformation. If you're a project manager, you've got university courses, certifications, Jira, Asana, decades of best practice to draw on. If you're an AI champion, you've got… a ChatGPT login and a lot of Googling.

    What strikes me most is how isolated these people are. Every AI champion I've spoken to is working it out alone, inside their own company, reinventing the wheel. They're not learning from each other. They're not sharing what's working. There's no community or infrastructure connecting the person figuring it out at a £10m consumer brand with the person doing the same thing at a £50m one.

    That feels like a massive missed opportunity, for the individuals and for the businesses betting on them.

    I don't have the answer yet. But I've been thinking a lot about what support for AI champions could actually look like, and I think it starts with getting these people in the same room.

  • The Imagination Gap

    AI is asking you to imagine something you've never seen. And it's hard to be what you can't see.

    I was speaking to the CEO who said 'we lack the imagination to even know what's possible', and it really stuck with me.

    Unless you're really into AI, it really is hard to see what to do beyond 'get ChatGPT licenses for everyone' because how are you supposed to know what's possible when very few people in this space are pushing the boundaries?

    The way I'm trying to bridge this 'imagination gap' for consumer brands is by taking inspiration from the businesses and the individuals who are operating right at the edge.

    Here are 3 sources of inspiration for me at the moment:

    'Companies as Code'
    If AI can do extraordinary things when given a set of 'skills' – essentially just well-structured text files – what happens when you codify your business in the same way? Your processes documented, your data accessible, your operating system written down in a form that AI can act on. So you don't just aim to be a consumer brand that uses AI, but a consumer brand whose operations are built to run like software.

    AI driven software development
    I spent time with a group of AI engineers who weren't just using AI to write code. They'd built entire systems where AI writes the code, tests it, catches its own mistakes, and iterates until it works. Their focus wasn't on building software. Instead, they'd built the machine that builds software. What's the equivalent for a non-technical business?

    See also – compound engineering.

    Claude Code for knowledge work
    Claude Code for non-technical use cases has been having a moment (and OpenClaw is Claude Code in a trench coat). If you want to really geek out, explore the Claude Code+Obsidian combo. Claude Code has become the default way in which I work and am building my consultancy business.

  • The Rise of the AI Operator

    Ops teams are always under pressure to cut costs, but they rarely get prioritised for dev time or allocated budget for new tools. They're stuck firefighting with no space to build something better.

    If you're an ops person who learns to build automations or AI tools – even small ones – you can create tools with massive leverage. And you're uniquely positioned to do this well. You're in the weeds so you see exactly what's causing drag and with AI, you can build something quickly to fix it (or at least make it less bad).

    The impact compounds as you move through the operation, continually improving it, fixing bottleneck after bottleneck.

    If you don't have these skills in your team, you're stuck doing everything manually, hiring linearly with growth.

    There's going to be a divergence where two similar businesses look identical from the outside, but one needs half the people to run the same operation.

    The role of 'AI operator' doesn't really exist yet as a job title. But it will. And the companies who figure out how to build or hire or train for it are going to have a structural cost advantage that's very hard to compete with.

  • Claude Code Plugin for Obsidian

    Last year I moved all my notes and files out of Notion and into Obsidian. With Notion, you're locked in to whatever tools they give you. I wanted to be file-first rather than software-first, so I could use the latest AI tools on my actual files.

    I was already using Claude Code to work with my files. Then I found a community plugin for Obsidian that adds Claude Code as a sidebar. I now have access to Claude Code and everything it can do right where my notes are.

    The Solution

    I installed the Claude Code plugin from Obsidian's community plugins. It adds a sidebar where I can have conversations with Claude Code whilst viewing my files. There's a button to link specific files into the conversation, so Claude has context on exactly what I'm working with.

    This isn't just having a conversation with Claude, this is Claude Code, which means it can take actions, plan, and has all the power that Claude Code has rather than just Claude, which has been way more useful.

    How I Use It

    As a practical example, I keep Markdown files for potential clients in my CRM folder. After a sales call, I use the sidebar to automatically pull in transcripts and add my personal notes and thoughts to create really high-quality notes about each client. It updates the metadata properties at the top of the file. That's how I track what needs to happen next – things like 'awaiting proposal' or 'follow up in two weeks'.

    I can also ask it, 'Who do I need to follow up with?' and it'll search my CRM notes and draft emails for each person.

    The whole thing takes a fraction of the time it used to, and to be honest, this kind of thing I often used to not be super on top of because it felt like a chore. Now I actually do it, or rather, now it gets done for me.

    Why This Has Impact

    Better follow-ups, less prep time before meetings, and I can easily search my notes to find who might be a good fit for something.

    The bigger win is flexibility. I'm not locked into any specific tool. My notes are just Markdown files. If a better AI tool comes out tomorrow, I can use it on the same files. File-first means I control my data and what I can do with it, instead of being locked in.