No Process, Just Prompts: Why Solo AI Projects Fall Apart

Morning. Coffee. You open the laptop and look at the notebook page where you sketched the plan three days ago. Arrows, boxes, a few crossed-out lines. You can't tell which box you're in. The agent can't either.

So you do what always works: you open a chat and start prompting. It feels productive by 10 a.m. By Friday there are three half-built features, two abandoned branches, and a plan nobody follows.

This is the first post in the series No Process — about why AI work without a process doesn't add up.

The short answer#

• Prompting without a process feels fast on day one and falls apart once you have several projects and several sessions.
• A minimum process for AI work has four parts: fixed phases, one next step per task, state kept in the repo, and a daily start and end routine.
• The process is not for the agent's sake; it's so you can stop and restart without losing the thread.
• Write the process down once, reuse it for every project of the same kind.

Why "just prompting" works — at first#

With one small project, prompting without a process is genuinely fine. Everything fits in your head and in one chat. The agent is fast, you're motivated, and there's no state to lose.

It breaks at three moments:

1. The second session. The new chat doesn't know what the old one did.
2. The second project. Now two sets of context compete for your head.
3. The first break. A weekend away and you can't remember where you stopped.

Each of these is where a process earns its keep. Not a heavy one — a small one.

The minimum process: four parts#

1. Fixed phases per kind of project. An app goes through the same steps every time: idea, spec, build, test, listing, release. A book: niche, outline, draft, edit, cover, publish. Write the list once and reuse it. This is the core of phases, not to-do lists.

2. One next step per task. Not a backlog of twenty items — the single next action. "Write captions for 3 screenshots", not "work on listing".

3. State in the repo. Phase, next step and key decisions live in a file next to the code. Not in your notebook, not in an old chat. Your agent reads it at the start and updates it at the end.

4. A start and end routine. Five minutes at the start: what's stuck, what moves today. Five minutes at the end: update the next step so tomorrow starts cleanly.

That's it. Everything else is optional.

What a process buys you#

The biggest win isn't speed. It's that you can stop without losing the thread.

• You can close the laptop mid-task and pick it up on Monday.
• You can start a new session and the agent continues instead of restarting.
• You can run several projects without them bleeding into each other.
• You can see at a glance what's done, what's next and what's stuck.

Without a process, all of that depends on your memory — which is how you end up working for your AI agent instead of the other way round.

Write the process once#

Don't invent the phases every project. Write one page per project type:

APP:   1 Idea · 2 Spec · 3 Build · 4 Test · 5 Store listing · 6 Release
GAME:  1 Concept · 2 Prototype · 3 Core loop · … · 14 Launch
BOOK:  1 Niche · 2 Outline · 3 Draft · … · 9 Publish

Put it in the repo. Ask your agent to follow it and to keep the phase up to date. The end-to-end AI workflow walks through one full cycle.

The rest of this series#

The next posts in No Process each take one symptom — forgetting the next step, losing a blocker in the list, coming back after a break, not remembering how to run a project — and show the small habit that fixes it.

Keep reading#

• The Solo Builder's AI Agent Workflow, From Idea to Release
• Phases, Not To-Do Lists: Project Management for Solo Devs
• Are You Working for Your AI Agent? The Copy-Paste Trap