A morning idea list that is not allowed to do anything

Every morning a small program on my home server reads through my notes and hands me a short list of things I could work on. On Wednesday it gave me five ideas: one for my studio software business, two about my career, one about the automation system itself, and one from the general tech and AI reading I had done that week. I read the list over coffee and closed it.

That is the whole interaction, and it is the part I want to write about. The tool suggests. It never does anything.

Why I built it

I have a lot of parallel tracks running. There is the day job, a software product with real customers, a master's project, a home server that runs a pile of automations, and the slow background work of figuring out what comes next in my career. Each of those tracks produces notes. Meeting notes, research summaries, half-formed plans, things I said I would come back to.

The problem was never a shortage of ideas. The problem was that good ideas got buried in the notes and I forgot they existed. I have ADHD, and for me anything I cannot see on a screen I check every day effectively does not exist. A great idea sitting in a markdown file I wrote a while back is the same as no idea at all.

So I wanted something that would read the notes for me, pull the threads out, and put a small ranked list in front of me every morning.

How it works

The digest runs early, before I am up. It gathers material from a fixed set of places: my project notes for the studio business, my career notes, the open questions in my server's architecture docs, and whatever research my weekly research job has pulled in recently. Each area gets its own pile of source material, with a pointer back to where each piece came from.

Then a model running locally on the server reads each pile and pulls out a handful of candidate ideas. I wanted the default run to cost nothing, so the heavy lifting happens on my own hardware rather than a paid API. Each candidate gets scored on two things: how much it would matter, and how hard it would be. The ranking rewards ideas that are high impact and low effort, which is exactly the kind of thing I tend to skip past when I am deep in something bigger.

Before anything makes the list, it gets checked against a running ledger of ideas I have already seen. If the same suggestion keeps coming back, it gets dropped. Nothing kills trust in a daily list faster than seeing the same item every morning.

The output is one page in my notes vault, dated, with each idea, why it matters now, and a concrete first step. That is it.

The rule that matters most

The design decision I care most about is that the digest is advisory only. It writes one page and appends to its own ledger. It cannot edit code, open pull requests, send email, or change anything in any of my projects.

This was deliberate. I have spent a lot of this year building automations that do act, and I have learned how much care that takes. Every tool that can change something needs guardrails, a way to verify what it did, and a way to undo it. An idea generator does not need any of that if it simply cannot act. The safest permission is the one you never grant.

There is a second guard I like. When the system proposes ideas about itself, it automatically throws out anything that would loosen its own budgets or limits, give itself more tools, or remove a safety check. That is checked in code, not left to the model's judgment. A system that suggests ways to improve itself will, sooner or later, suggest giving itself more room. I would rather that suggestion never reach me than rely on myself to say no every time.

What it is actually like to use

Most mornings I read the list and do nothing with it. That is fine. The value is not that every idea is good. The value is that the good ones surface at all, and that I see them on my schedule instead of stumbling on them weeks later.

Wednesday's split was a fair picture of where my head is right now. Two of the five ideas were about career, which matches how much of my thinking has gone there lately. The single studio idea was small and practical. The tech idea came straight out of reading I had already done and half forgotten.

What is still open

The scoring is the weakest part. A local model rating its own suggestions for impact and difficulty is a rough signal, and I still override it in my head most days. I also want a cleaner way to mark an idea as done or rejected, so the ledger knows the difference between "seen" and "acted on."

But the core idea has held up. A tool that reads everything, suggests a little, and never touches anything turns out to be one of the most useful things I run.