The OS
It's staggering how far AI has come, but LLMs are still apps missing an OS. For me it started with opening Claude every morning and having to re-explain everything we'd been working on — what fund, where we'd got to, what decisions had been made. Claude can remember general things about you if you've set that up, but it has no idea what we were working on yesterday. That operational context? You always have to manage that yourself.
Once I had the OS running I realised it was solving a bunch of things at once — context window limits, accumulated knowledge across funds, continuity across multiple threads. People tell me about different walls they've hit, but the fix is the same.
↳ So here's what I built.
How I got here
If AI is so smart, why can't it be more like a team member I'm training rather than a chat box? You know, like learning what it's been taught so it doesn't have to come in every day and start from scratch.
After a lot of learning and trial and error, it took three files, a habit, and a bit of overnight automation to get closer to what I thought AI should be able to do.
I'd have a great session working through a fund, get somewhere useful, close the tab — come back the next morning and it knew nothing. Had to re-explain everything. Portfolio, mandate, where I'd got to. Twenty minutes of context just to get back to where I was.
Claude Projects helped a bit. But it only holds static background — it doesn't know what I was doing yesterday, what decisions I made, what's in progress. So I built that myself. Three files. Turns out that's all you need. This works whether you're just using an LLM or gone for the full AI stack. The files are the same either way.
Setting up the files
Before anything else, you need somewhere to keep your memories and operating guidance in files. I use a folder on my desktop called investment-os/ but it genuinely doesn't matter where — Notion, Google Drive, iCloud, a folder on your desktop. Anywhere you can open a text file quickly.
Create a folder. Inside it, create a plain text file called working-context.md — or working-context.txt if that's easier. The format doesn't matter, just the content. (As an aside, try to get used to md files and Obsidian — they just hold text and are a fraction of the size of .doc files.) You can do this in any text editor — Notes on Mac, Notepad on Windows, Notion, Obsidian, whatever you use.
Not sure what to put in it? The easiest way to start: open Claude, paste the template below, and tell it a bit about your work. Ask it to fill in the file for you. You'll have a first version in two minutes.
# Working Context Last updated: [date] ## What I'm working on [One sentence — the active task or decision right now] ## Where I am [Progress so far, key findings, decisions made] ## What's next - [ ] Step 1 - [ ] Step 2 ## Open questions [Anything unresolved you don't want to lose]
Save the file. That's it. You can add hot-context.md and the topics folder later once the habit is in place.
Three files. That's it.
Folder on my desktop called investment-os/. Three files inside it. Nothing clever.
investment-os/
├── working-context.md # what I'm working on right now
├── hot-context.md # what happened in the last day or two
└── topics/ # one file per fund or subject I keep coming back to
├── private-credit.md
├── infrastructure.md
└── [fund-name].mdI started with just working-context.md. Didn't even bother with the others for the first week. Get the habit first.
Starting a session
I use Claude because it seems to be the most capable LLM — particularly for long context work. But it doesn't matter what model you're using. This will change the way you work with AI regardless. If you tell me another model is performing better than Claude, I'm with you.
Open a new Claude chat. Before you type anything else:
1. Open working-context.md — select all, copy, paste it into the chat 2. Do the same with hot-context.md if you have it 3. If you're deep on a specific fund or subject, paste that topic file too 4. Then just get on with it — pick up the DD, ask your question, whatever you came to do. No preamble needed.
Two minutes. The AI is fully up to speed. Picks up exactly where you left off. The files are yours, not the model's — works with any LLM.
Ending a session
Before I close the tab I run the prompt below. It tells Claude to update all three files based on what we just did.
Copy the output, save it in the relevant file — overwriting the version you started the day with. Done. Tomorrow's starting point is already updated.
Feels like overhead the first few times. After a week it's just part of closing the laptop.
Before we finish, please update my context files: 1. working-context.md — where I am now, what's next 2. hot-context.md — quick bullets on what happened today 3. [topic file if relevant] — anything new worth keeping Output each in full so I can copy and save.
This is the manual night shift. Not automated — just a prompt and some copy-paste. The habit is what makes the whole thing work.
The night shift
The manual version above is genuinely useful. But there are three levels — and the jump from Level 1 to Level 2 is a real step, not just a setting you flip. I'm just giving you visibility into the road I've travelled, but it may not even be possible unless you have an AI strategy that allows agents to operate in a silo like external contractors.
Level 1 — just Claude. Manual paste at the start, manual prompt at the end. Five minutes total. Works today with nothing extra. This is where most people should start.
Level 2 — you introduce local tooling on your device that sits alongside Claude. It reads your context files automatically at session start and writes them back at the end. No need to paste anything — you just open a session and get on with it. But this requires setting up something beyond a Claude subscription like OpenClaw, which isn't secure enough to load onto your main work laptop. I've done it on a separate device as this is where things are going, but it's a meaningful step up, not a quick toggle.
One thing worth flagging: any tooling that sits on your machine and accesses your files raises legitimate security questions. No institution will approve this kind of setup today, and even for personal use it deserves thought. The way I've handled it — a completely separate machine with zero credentials, zero access to any sensitive accounts or systems. Like an outsourced offshore contractor — they get access to what they need to do the job, nothing more. No keys to the filing cabinet.
Level 3 is what I run now — everything happens overnight. Context files updated, data refreshed, morning briefing ready when I wake up. The AI starts each day already knowing what happened yesterday. This is where it gets genuinely interesting.
Start at Level 1. It works.
Checkpoints
For anything that spans multiple sessions — deep DD, building a model, working through a long DDQ — I get Claude to write a checkpoint file before we start. Just a quick note of where I am and what's left so we can hit the ground running at any point.
If the session dies mid-task, Claude reads the checkpoint first and we pick up straight away. Never lost real progress since I started doing this — particularly useful when you get more advanced and start having an LLM build things that take a while to complete.
# WORK IN PROGRESS — [Task Name] Started: [date] ## What I'm doing DD on [fund] ahead of IC ## Where I am - Factsheet and quarterly letter reviewed ✓ - Key figures extracted ✓ - Comparing yield vs hurdle rate — in progress ## What's next - [ ] NII coverage ratio - [ ] Leverage vs peers - [ ] IC summary draft ## Notes Fee restructure mentioned in Q3 letter — flag this
The queue
At Level 1 this is just a well-organised to-do list. Claude has no ability to work through it on its own — it can't wake itself up, pull the next task, or run overnight. That's a Level 2/3 thing.
What it can do is give you discipline. I keep a queue file — one list of everything that needs doing across active research threads. I review it at the start of every session, pick one thing, and tell Claude what we're working on today. When it's done I mark it complete and pick the next one.
Simple but it stops things disappearing. When you're tracking five or six funds simultaneously you need something to hold the thread — otherwise things fall through the cracks between sessions.
At the full stack level is where it gets really powerful — the queue runs sequentially and autonomously overnight. Claude isn't a great multi-tasker, so agents work through it one task at a time while I'm not there, mark tasks complete, log what they did. The file format is the same. The autonomy is completely different.
## This session - [ ] Fund X quarterly letter — check NII coverage - [ ] Hurdle rate vs current gilts — update the numbers ## This week - [ ] New infra fund — initial screen - [ ] IC memo for private credit position ## Backlog - [ ] Sector exposure analysis
Topic files
One file per fund or subject I keep coming back to. Load it when I need it, ignore it when I don't. Keeps the context lean.
Mine look roughly like this:
# [Fund Name] Last updated: [date] Strategy: European direct lending, senior secured Stage: in DD ## Key facts - Target return: net 8-10% IRR - Fee structure: 1.5% / 15% carry - Liquidity: 3 year lock-up, quarterly thereafter ## Current view Interesting entry point. Yield clears the hurdle at current gilts. NII coverage looks fine but worth double-checking Q3. ## Open questions - [ ] NII coverage ratio — Q3 letter - [ ] Co-invest rights — ask at next call ## Log - Apr 14 — Initial intro, passed hurdle rate screen - Apr 21 — Deeper DD started
I don't have one of these for every fund I've ever looked at. Just the ones I'm actively tracking. Grows as you need it.
Getting help
If any of this feels unclear or you want to adapt it for your own setup — just ask Claude.
Paste your working-context.md (even a rough first version), describe what you're trying to do, and ask it to help you build the system. It's surprisingly good at this. I've had it help me restructure my context files, suggest what to track in topic files, and debug why something wasn't working.
You don't need to figure it out alone. That's kind of the whole point.
I'm trying to build a context system for my investment research using Claude. Here's my current working-context.md: [paste your file] I want to [describe what you're trying to set up or fix]. Can you help me improve this?
My setup
For what it's worth, here's exactly what I'm running.
The machine: dedicated 32GB Mac Studio, 500GB drive — separate from everything, no credentials, no access to anything sensitive The models: - Claude Sonnet — analysis, research, judgment calls - Local open-source models — bulk tasks, document ingestion, overnight monitoring (free) The setup: - Three context files (working-context.md, hot-context.md, topics/) - Task queue that runs overnight - Knowledge wiki — fund docs, research, articles, all ingested and queryable - Specialist agents for research, data, writing Total cost: Claude Pro + hardware. Local models are free.
Most people will get a huge amount of value from Level 1 — just the three files and a Claude subscription. But this is where I ended up after a while building, and it works well enough that I wanted to document it. And it is where the puck is going.