Track: planning and running projects
The honest summary of this track: AI is very good at two specific parts of planning and actively harmful at a third.
It is good at getting the thing out of your head — turning a vague intention into something written down with the gaps visible. It is good at keeping a plan current — reconciling what you said you would do with what you actually did. It is bad at deciding what matters, and it will produce a beautiful document that helps you avoid deciding.
Know which of those three you are doing and this track is short.
---
Stage 1 — getting it out of your head
You have a thing you intend to do and it exists only as a feeling.
What to try first. Say it out loud, badly, in one paragraph. Then:
- "What am I assuming here that might not be true?"
- "What is the first thing that has to happen before anything else can?"
- "What would make this take three times longer than I think?"
- "What have I not mentioned that somebody doing this would need?"
You are not asking for a plan. You are asking for the questions you would have been asked by somebody experienced, which is the part that is genuinely hard to do alone.
What goes wrong. Asking for a project plan and getting one. It will be well-formatted, generically sensible, and about a project that is not yours. The tell is that it contains phases called Discovery and Implementation.
When you have got it. You use it to find holes, not to fill them.
---
Stage 2 — turning it into steps
You know roughly what needs doing.
What to try first. Give it your rough list and ask for one thing at a time:
- "Which of these depends on which? Just the dependencies."
- "Which of these is actually two tasks?"
- "Which of these could I not start tomorrow morning, and what is missing?"
That last question is the useful one. A task you cannot start is not a task, it is a wish, and it is the commonest reason plans stall.
What goes wrong. Accepting estimates. It has no idea how long anything takes for you, and a plausible number is worse than no number because you will half-believe it.
When you have got it. Your lists are made of things you could start, and you know what each one is waiting on.
---
Stage 3 — keeping it honest
The plan exists and reality has started disagreeing with it.
This is where most planning tools fail and where AI is genuinely, unglamorously good — because the job is reconciling two texts, which is exactly what a language model does well.
What to try first. Give it what you planned and what actually happened — notes, commits, a calendar, a journal, anything real — and ask:
- "What did I say I would do that I have not mentioned since?"
- "What did I actually spend time on that was not in the plan?"
- "Which of these is finished and which have I just stopped talking about?"
That last distinction is worth the whole exercise. Every stalled project has three or four items that were quietly abandoned without ever being decided against.
What goes wrong. Not giving it the real material. A model asked "how is my project going" with nothing to go on will produce encouragement.
When you have got it. You do this weekly and it takes five minutes.
This is the specific job AIOS was built for: you write the day down normally, and the plan, the calendar and the daily brief are read out of what you wrote rather than maintained separately.
---
Stage 4 — the routine
The planning stops being a project and becomes a rhythm.
What to try first. One review, at a fixed time, with a fixed set of questions. Weekly is right for most people. The questions matter less than them being the same questions every time, because that is what makes the answers comparable.
A serviceable set: what moved, what did not and why, what is now blocked, what should I stop doing.
What goes wrong. The review becoming a document nobody reads, including you. If it takes longer than about ten minutes it will not survive a busy week, and a routine that only runs in calm weeks is not a routine.
When you have got it. You would notice if you missed one.
---
The tools
Whatever holds your notes matters more than any planning tool. The stage-3 jump needs your actual week to exist somewhere in writing. A journal, meeting notes, commit messages, a calendar — real traces beat a well-maintained board.
A chat interface does everything on this page. The context you give it is the whole game; the model choice barely matters.
A project tool if you already use one. Most now have AI features. They are good at summarising a board and bad at deciding what should be on it, which is the same split as everything above.
AIOS, if the reconciliation part is what you want and you would rather not do it by hand.
---
Where it genuinely does not help
Priorities. It will rank your tasks. The ranking will be reasonable and it will not be yours, because it does not know what you are actually trying to achieve this quarter or which of your commitments you privately intend to drop.
Estimates. Covered above and worth repeating.
Anything political. Who should own this, whether to have the difficult conversation, whether this project should exist. It will give you a balanced answer that is useless precisely because it is balanced.
Making you do the work. A plan is not progress. It is very easy to spend a satisfying hour producing an excellent plan and call that a day's work, and AI makes that hour much more productive-feeling.
---
Already written up here
- Time and Project Management for Programmers — the same problem without the AI, which is worth reading first.
- Steps Become Routines, Routines Become Systems — what stage 4 is actually for.
- Building a Routine That Survives a Bad Week — the failure mode every review process eventually meets.
---
Next: Everyday admin covers the smaller repeated jobs. Or back to all the tracks.