‘Todo apps’ feel to me like overhead, not support. This summer, I had a look at what’s out there with AI, but it was still underwhelming. Rather than another premium subscription, I started experimenting with building my own productivity system, focused on problems that were most important to me personally: the emotional labor of prioritisation, staying intentional with how I spend my time, and simply having a good record of my tasks over time.

Here’s a look at what I tried, and some lessons learned along the way.

“The fast fashion era of SaaS” (Sam Altman)

One hypothesis is that if average-quality software or features are cheaper, we’ll get a long tail of tools proliferating, and evolving rapidly.

It’s clear to me that personalized versions of your paid-for SaaS are very achievable. The trade off between fewer features and better alignment genuinely feels magical. So my project has a UI, but doens’t have WatchOS reminders. You might find todo.txt files and claude code completely cover your use case (like e.g. Teresa Torres’s approach).

Here’s a screenshot of what that part looks like today:

There’s one important caveat: I’m building alone, for myself, to get more done but also as a way to explore and learn about what’s possible.

“The tool should get out of your way” (Karri Saarinen)

As I added features, incremental building and the overhead of task management was getting more laborious. That shifted my perspective, and I started asking how much of the overhead I could offload.

Run locally for peace of mind

I started thinking about how LLMs could augment my planning, and I introduced this as a key constraint, because:

1. Credit anxiety is real on an llm-based service, or if you trust a coding agent with a paid API key. This influences the design, especially for things that should feel like a background hum, like categorising tasks or updating short descriptions.
2. AI shouldn’t replace my thinking. A lot of my tasks include some form of problem solving, and commercial LLMs tend towards average solutions. Being deliberate about the models, prompts, and agent design gives me a sense of control. I think this approach compounds over time. (see also, why I’m writing)

Here’s an early version, based on the idea that I want to make sure I’m spending my time intentionally (ie. avoid under-indexing on health and learning, really).

Detail deep-dive: don’t make me work

Having deep empathy with myself as a user, I’ve removed any attribute editing from the UI. That removes the temptation to be a super user and just get the boring bits done, creating a forcing function to fix the prompts, if there are issues.

One of the biggest wins here has been learning evaluations from the ground up. I started by rolling my own, with ChatGPT’s help, and it’s been a very helpful way to build up intuition for model performance and prompt-engineering.

“It’s not prioritisation till it hurts” (Ami Vora)

In all of this, there’s been one really big opportunity - going from task-tracking, to an assistant. Having connected my task database and locally running LLM, I’ve been looking at agent workflows. Simple RAG already has some “wow” moments for me.

Immediate wins - what would I do if I was in flow?

Here’s an example. When I’m in a state of flow, I find it easy to break down a task, progress, re-plan, etc. But if I’m between calls, or starting work on a lazy day, it can take me a while to warm up. I’ve really like having Giskard as a companion that channels the energy of my best self.

(me) I’ve got 30mn, pick a task for me

(giskard) Okay, I’ve fetched 22 tasks for you! There are a few learning-focused options, like “Evaluating AI Agents course” or “Discover how people are working through todos today?”. There’s also a shorter “Short Run (careful with knee)” if you’re looking for something more active.

Given you only have 30 minutes, I’d recommend “Evaluating AI Agents course” – it seems like a great way to spend a focused 30 minutes learning. Would you like me to go ahead and add that to your to-do list?

The prioritisation pain is real, and sometimes that can lead to procrastination. It feels a bit like the difference between a creative and editing mindset. It’s been nice to offload some of that emotional labour.

There are non-trivial limitations to running my own local models, especially because it feels like a bunch of extra scope which can slow down iteration. On the other hand, I am looking forward to applying some of the research on getting the most out of these smaller models.

Shout if you disagree!

This is an early snapshot of what I’m thinking. There’s a lot I’m excited about discovering - if you disagree or have any other recommended readings, please do let me know :)