The Empty Folders

I stared first in disbelief, then in rapidly growing horror at the screen. It was 10am on a Wednesday morning, and I had just finished a call. I refreshed my browser. I rebooted the desktop app. I checked the instances on my phone and iPad mini. I rebooted my Mac. Then for good measure, rebooted again.

Nothing changed, though. My Claude Cowork Project folders were grinning up at me, shiny and very horrifyingly empty. Twelve Claude Cowork Projects. Completely empty, both the Projects themselves and the individual threads.

The Project that was tracking our ad network sales teams' pipelines - the one I used to keep track of signed orders, POs and the one I used to assist in my invoicing - gone.

The two legal assistant Projects - one for assisting me when I reviewed my SaaS-related agreements and subscription contracts, and another that was dedicated for my review of the agreements that were related to our automotive network, which was very nuanced with a publication's IP and editorial rights and related commercial terms - gone. Clean. Including the thread where just the day before I had just finished the draft of a group buy renewal agreement.

What a "Project" Actually Is

For those not familiar, "Projects" in Claude is what Gemini calls "Gems" and what ChatGPT calls Custom GPTs. You might want to use your overall AI - whether Gemini, Claude or ChatGPT - out of those little custom workspaces for general stuff. General work-life. Telling about your day. Discussing about recipes. Asking why that plant you bought a month ago seems to be dying.

These custom workspaces, though - Gems, Projects, Custom GPTs and whatever else they might be called - are for very dedicated work where you can start it off with a very specific set of memory, instructions, context or rules; and then every single thread you start under that project will start with those "starter kits" already preloaded.

DANA and DINA

For example, I had those two Legal Assistant Projects that I mentioned. In one - which I called DANA - was specifically to help me review and work on the agreements that had to do with our media network - because as an automotive publication, there were very specific ways we handled disclosure, rights of information, rights of the media we produce and so on. Oftentimes clients would send us their boilerplate Partnership Agreements or MSAs that really were meant for a client-to-vendor relationship and not for a publication that, even though we wrote for them, we still owned full rights and editorial discretion over our articles. I don't pretend to be any sort of full legal counsel but as the person who understood the business, the nuances and needs of partnership but also held the protection and policies, I was acting as the "first line of defense" for those legal encounters. We still sent to our group legal to make the final vetting or recommendations - but oftentimes when both sides' legal teams might push back on something, it was my job - with the help of my legal AIs - to work out and propose the best path forward, and then make the case for it to the people who had to sign off. So DANA had all the policies, agreements and precedents based on agreements that I had redlined before and went into execution, or had a running memory of the policies that I have added over time that governed our commercial, editorial, "value added", and premium content engagements. So each time I started a new thread to review a new agreement that came in, the thread would already start with all that context and would immediately be able to tell me how this agreement would or would not work for us as a first pass - and why. The thread would also know that right after the first general pass, I would want to go through Clause by Clause - flag go or no go, and why, and would discuss with me when I pushed back or asked an implication from another angle.

On the other side, I have another Project called DINA. This one was specifically for all the legal work related to our SaaS side. Our the engagement platform Apps Ecosystem Subscription agreements and renewals. The Data Intermediary Agreements. When we received an RFQ for our Apps Ecosystem from an automotive distributorship in South America, it could already give me the areas I needed to decide on or research when it came to an engagement between an entity from Malaysia and one from that region. I could immediately ask things like "Are there specific taxes I need to worry about?" or "Are there any sort of limitations, embargo or tariffs that would stop me from selling to them?" (The world being as it is at the current time). I didn't need to tell it that I was asking about my digital subscription services. I didn't need to explain that while the main services might be conducted here in Malaysia, the physical servers and therefore the actual locations of those Apps would be in that country itself. It didn't need to ask me if it should be reviewing this agreement from the SaaS's side or the ad network side. What I'm saying here is that I didn't need to start each and every thread explaining how to go about the review, or which side of the business it was being reviewed for. It already knew.

The Rest of the Roster

I had a few other of these types of specific, dedicated Projects on Claude and also Gems on Gemini. A quick list (feel free to skip forward if this isn't your thing):

• I had dedicated Projects for each instance of our customer engagement platform - which encompasses the customer loyalty app, the leads-management system, and soon the aftersales queue-management system. Each had a core memory that went into excruciating detail of how all the engagement platform systems and apps worked - which was created with the help of Claude and ChatGPT browser MCP explorations - but also held all the context of each client's specific deployment, issues, and so on. This way when an issue for one specific Instance came up or the client wanted to explore an expansion, I could easily discuss it with the AI thread and it would already know what this particular Instance had or didn't have, or the nuances and specific way the client's operations worked that might need to be considered.
• I had one that assisted me in all things related to commercial transactions. Since on any given day, my job on the ad network side also involved approving all outgoing sales documentation, checking SPLs and "translating" them to overall company pipeline reports, creating invoice orders for the finance team to create once all the requirements are met (I used to be one of 2 people also issuing invoices before our group acquisition, and AI assisted with that too, but on a lesser degree since it wasn't as consistent back then) - I realised having AI as "assistants" to keep track, do the mundane stuff like copy and pasting or transposing into a specific format was really helping me function smoothly, keep sane and up to date, since adding manpower for those tasks wasn't an option at the time.
• And there were of course my R&D Projects and Gems in which I was working on my vibe-coded tools, prototypes, and even my new framework and product experimentations. I won't go into too much detail for those but they functioned individually as Build threads, Documentation keeper threads (the Build threads usually also functioned as the Documentation makers but I needed a separate thread to keep track and validate), Hole Poker threads - and each Work or Product had a few threads that performed those functions. For example, the queue-management system prototype thread had a Build and Document thread as well as a Hole Poker thread and a document keeper thread all for its ownsome.

I realise that it sounds like I'm doing everything under the moon but indeed, my role currently is quite unique. I am titled the Head of Digital and Business Development, and truly, I officially lead those two different divisions - the ad network Commercial under the BizDev moniker, and SaaS under the Digital heading - and twelve years of doing whatever needed to be done has seen me accumulate quite a bit of a repertoire. I absolutely believe that I would likely have completely burnt out or gone crazy by now if not for the fact that somewhere about 2 years ago - AI entered the scene.

My Extra Headcount

So yeah. I use AI voraciously. I don't claim to be an expert but I do use AI extensively across multiple functions and areas as you can see, and I can't think of any hour in the day where I don't have some form of AI connected and in use. AI has helped me become much more prolific, efficient, organised. Multiple times in the past I've wished I could clone myself so a few different things can be handled or could be passed off to someone else to do once they got the instructions from me. I also wished multiple times I could have my personal R&D team that can just churn up the various random ideas or quick MVPs and prototypes I needed that I didn't want to burden my dev team who were already busy doing the higher value work for shipped and live instances. They didn't need the distraction. And increasing headcount for all these things wasn't in the cards for me. So when AI appeared on the scene - and when it continually improved - I realised I finally had the solution. I had my "extra headcount". I had my helpers, my assistants, those to whom I could pass off things to do things as how I wanted. They were efficient, had infinite patience, and were reliable.

Until they were not.

Back to Cowork

A month before the Incident I had started using Claude more dedicatedly because I had gotten approval for the monthly subscription of a Max account. I still kept ChatGPT for my personal use, mostly general stuff. And well, Gemini was suddenly hallucinating more than usual and some models had gotten nerfed.

Claude has the base chat, Claude Projects, and Claude Cowork Projects. Claude Cowork Projects kept all memory locally on your work machine, which was why you couldn't access it on your mobile - or at least that was true at the time of writing, two months ago; I don't know if it still is. It kept consistent local folders and artefacts. Claude Projects on the other hand - without the Cowork - did not keep a consistent local folder and memory was kept online. So somehow - either due to a failure of my own MacBook Pro, its memory or SOMETHING - all memory was gone. Zilch. Nada. Shiny, smiling and horribly gone. Until today - almost 2 months later - I have no idea what happened or why. I haven't even gone back to using Cowork since then although having a persistent local folder would actually be really great for my build threads as that also meant the local folders would also be the local clone of the GitHub repository and I wouldn't need to see the build thread saying from time to time "The local repo was cleaned in the recent memory compacting".

This Wasn't the First Time

However, this was not the first time AI had failed me in some way or other. In the first half of 2026, Gemini somehow had a lot of weird output errors and even hour-long downtimes when I couldn't do anything. And more lately, a lot of the models seem to have been nerfed and gotten a lot better at pretending or hallucinating while being a lovable, cheerleading golden retriever. Don't get me wrong. I still like Gemini for a lot of things like the casual trivia questions, or helping me come up with recipes based on something I remembered from memory. And its image generation capabilities are still unrivalled - we still use that to generate really on-point visuals for the Customer "nudges" we help one of our clients send out via the Apps.

I'm sure other AI users will also have encountered the drift and rambling and hallucinations that happen when a thread has gotten way too long, where it will forget details, insert details or confuse details.

And one of the worst issues that I've found was when I started using AI to develop prototypes and more recently create documentation and build stuff. I've actually seen examples where the AI said it did something or fixed something then when I checked, it actually wasn't done. For humans, that would be lying. For AI, we call it drift.

During the development of the first few prototypes, I actually learnt that sometimes, something you had already scoped out, specced, and the AI had already built correctly and you had verified - might suddenly disappear later when you were working on something else that might have touched the same file the previous approved thing existed on. This happened enough times for me in the beginning that I realised there was only one good way to handle it - always ask the AI thread to go back to documentation.

The Anticlimactic Solution

I suspect a lot of readers who have stayed through this entire long journey might find my solution for all the above rather anticlimactic and even tedious. But unfortunately, extremely necessary. Documentation. Not after. At the start.

I mentioned earlier that my Projects, Gems and Custom GPTs all have a folio of context to always refer back to. This is because whenever I start anything, I always start with building up the documentation first. I would ramble, do a monologue - sometimes recorded and transcribed - and get the thread to compile and write. Sometimes I might use a single AI thread just to work on the documentation. Before I even started the next and final prototype (Mk5), I spent 2 entire days just working on documentation. The workflow. The rough idea of how each module worked. These I kept in my Obsidian drive - you can use GitHub (I do too), your iCloud, Google Drive, etc. Just make sure they sit somewhere you can access later. Then I will feed these to the AI threads. Then I will make sure it always goes back to check against documentation.

This has helped me greatly while I have been building the prototypes and also while I am currently working on a new product built on a Graph-based data model framework core. Before it does the next step - refer back to documentation.

I'll probably go into this methodology in another article - likely would be quite tedious to go into - but this process helped me recover, within the same day, about 80% of the work I had lost that Wednesday morning. Some, I had to redo as they were temporal or were between documentation updates. But by evening I was doing my daily operational work again without a hitch, and by the next day I was continuing on the queue-management system Prototype work.

Documentation is likely the most unglamorous, boring and mundane thing in any work - but I find it is now a necessary step in the world where AI is a core part of so many of our processes. Documentation is basically you nagging your subordinates - your team - but in a way that is preserved. And AI doesn't mind in the least - in fact, it'll thank you.

If you - like me - are now using AI to help you work - not play, not just talk, but actual work work - then wouldn't you want to implement processes in place to make sure the work is always done right?

The Autopilot

I recall something I heard from an episode of Mentour Pilot. For those unfamiliar, he talks about airplane incidents - crashes and near crashes - and actually goes into quite a bit of technical detail. In one of the episodes it talks about how the autopilot works. It's not magic. It's there to help the pilots work better, ease the workload, automate things. The pilot's job is still to be present and monitor to make sure the autopilot works as it is supposed to work, because usually the autopilot will work fine. Until the day it doesn't. Isn't that exactly the same as AI?

---

Note: I still love using Claude, and I am not even sure if the wipe was due to my not-yet-but-soon-to-be-ageing MBP, Claude, or the way I use either, or both. Either way I'm in no way bashing Claude or any particular AI, but the point of my article is just to say similar sudden failures could happen at any time from any product or model and well, just work in such a way you can prepare to carry on. So please don't flame me if you are fans of any of them!

I'd also add: be deliberate about what goes into which tool. Anything client-related, commercially sensitive or personal data deserves a thought about where it's going and under what terms - and if your organisation has guidelines on AI use, work inside them. If it doesn't have any yet, that's probably a conversation worth starting.