METR measured it in 2025 — senior devs with AI coding assistants worked 19% slower and thought they were 43% faster. For side projects, this gap is career-ending for your product.
In July 2025, METR ran a controlled trial with experienced software developers using AI coding assistants.
The result: they worked 19% slower than without AI tools. Not faster.
Worse: they thought they were working 43% faster.
A 62-percentage-point gap between perception and reality.
I kept thinking about what this means for side projects specifically — because I think it matters more there than anywhere else.
The planning fallacy didn't go away
Daniel Kahneman named it in 1979: we systematically underestimate how long our own tasks take, and overestimate how much we'll get done. External reference points help. Deadlines help. Someone watching helps.
The assumption was: AI will fix this. AI will plan better, build faster, reduce rework.
METR's data says otherwise. Not because AI tools are bad — they're genuinely capable. But because the planning fallacy is a cognitive pattern, not a speed problem. The issue isn't how fast you can execute a decision. It's whether the decision was right in the first place.
What actually changed
AI tools like Claude Code, Cursor, and Codex shifted something real. The problem isn't "I can't build this feature." You can build almost anything faster than two years ago.
The problem shifted:
"I'm building very fast in the wrong direction."
Without structure, every session starts with: what do I actually work on today? With AI, you can execute that decision faster — but if the decision is wrong, you've just done more damage more efficiently.
For a solo dev with a full-time job, this compounds. You have 90 minutes on a Tuesday evening. You fire up Claude Code. It helps you build. But toward what? The commitment is to the project in your head — not to the specific thing that makes a user pay.
The side project-specific problem
Side projects don't have a PM. No sprint planning. No "why are we building this?" meeting that forces articulation. No one asking "is this the right thing to ship this week?"
For a full-time team, structure is often annoying overhead. For a solo dev with 8 hours a week, it's the difference between shipping in three months or archiving in six.
AI makes you more productive within a session. It doesn't know that you skipped Day 4. It doesn't know that you spent your last three sessions on a feature nobody asked for. It won't ask you "is this the thing that makes someone pay?"
That accountability — the external forcing function — has to come from somewhere else.
What I built
I ran into this loop three times across three different products. I started building a system to break it.
MVP Builder is a structured 30-day sprint with one daily AI prompt, delivered to your inbox each morning. The prompt isn't generic. It knows your project — your stack, what's built, what's left, what you said yesterday — and asks you to move one specific thing forward.
Three tiers, based on where you actually are:
- Bronze (13 days): Idea only → working prototype
- Silver (21 days): Started but stuck → shippable product
- Gold (30 days): Almost done → actually shipped
The structure is the feature. The daily prompt is the external "what do I work on today?" so you don't spend your 90 minutes answering that question yourself.
Where I am now
Cohort #1 is free. Beta. 5–8 spots. I'm in Day 10 of recruiting.
The product works — the cron system, the prompt pipeline, the milestone tracking. The part I'm still figuring out is distribution. Which is its own kind of side project problem.
If you've shipped something after a long "almost done" phase — or if you're currently in one — I'd genuinely like to know what finally broke the loop.
Cohort #1 is free: mvpbuilder.io/pipeline
Building in public. Previous post: I built 3 MVPs that never shipped. Here's what I learned.
Top comments (0)