100,000 New AI Projects a Day: Why Most Stall After the First Draft — and How to Brief Yours
More things are being started with AI than ever before — one app builder alone reports more than 100,000 new projects every day, and GitHub saw about 121 million new repositories in 2025. Most of those projects never get past the first impressive draft. The usual reason isn't the tool or the model: it's that the AI was given a one-line wish instead of a brief. This article shows the numbers, why projects stall, and the short brief that keeps an AI-built project moving.
How much is being built
App builders. Lovable, one of the best-known "vibe coding" tools, said in July 2025 that more than 10 million projects had been created on it, with over 100,000 new projects started every day, and repeated the 100,000-a-day figure in its one-year post in November 2025 (Lovable, One year of Lovable). Earlier in 2025 that figure was closer to 25,000 a day (analysis on Substack). That is one product among many.
Code. GitHub's Octoverse 2025 report counted about 121 million new repositories in 2025 — roughly 230 a minute — and 36 million new developers, about one new developer every second. It counted 4.3 million AI-related repositories, and more than 1.1 million public repositories using an LLM SDK. Nearly 80% of new developers used Copilot in their first week (GitHub, Octoverse 2025).
People who never coded before. A large share of those builders are not engineers. They are shop owners building a booking page, teachers building a quiz app, founders testing an idea over a weekend. AI made the first version cheap. It did not make the right version automatic.
Why most of it stalls
If you have built anything with an AI app builder or a coding assistant, you know the curve: the first ten minutes feel like magic, the next ten hours feel like arguing with someone who keeps forgetting what you agreed.
That pattern matches what researchers see in companies. MIT NANDA's 2025 study of enterprise AI described the core problem as a "learning gap" — AI set-ups that don't keep context, don't remember feedback and aren't fitted to the actual workflow (Virtualization Review). The same thing happens on a one-person project. Four things usually go wrong:
- The goal lives only in your head. "Build me a booking app" leaves hundreds of decisions to the model: who books, what they book, what happens on cancellation, which payment provider, what the admin sees. It makes all of them for you — differently each time.
- There is no definition of done. Without acceptance criteria, every change is judged by "does it look right?", so fixes break things that already worked.
- The context resets. New chat, new session, new agent run — and the decisions from yesterday are gone unless something writes them down.
- Big asks, no steps. "Add user accounts and payments" in one go produces a tangle. The same work in five small, testable steps produces something you can keep.
The brief that keeps a project moving
Before you open the builder, spend fifteen minutes on a one-page brief. Paste it at the start of every session, or save it as a project file your AI tool reads automatically.
1. One sentence of purpose. "A booking page for my two-chair barbershop so regulars can book and pay a deposit without calling."
2. Who uses it, and what each person must be able to do. "Customer: see free slots for the next 14 days, book one, pay a €5 deposit, cancel up to 12 hours before. Me: see today's bookings on my phone, block out time, mark no-shows."
3. What it must never do. "Never take full payment. Never show other customers' names. Never send more than one reminder."
4. Decisions already made. Stack, payment provider, colours, language, where it is hosted. Each one you write down is one the model stops re-deciding.
5. Definition of done for the current step. "Done when: a customer can book a slot on a phone, the slot disappears for others, and I get an email." One step at a time.
6. A running decisions log. After each session, ask the AI: "Summarise what we decided and changed today in five bullet points." Add them to the brief. This is the memory most set-ups lack.
That's the same five-part shape we use for every prompt — Role, Context, Task, Constraints, Output — scaled up to a project. The RCTCO guide shows it on single tasks.
Prompts that work well on builders and coding agents
These short instructions consistently produce steadier results:
- "Before writing any code, list the files you'll change and why. Wait for my OK." It stops sprawling rewrites.
- "Write a failing test that reproduces the bug first, then the smallest fix that makes it pass." It turns guessing into a check.
- "Do only step 2 from the brief. Don't touch anything else." Scope protects what already works.
- "What did you assume that I didn't tell you?" It surfaces the silent decisions before they become bugs.
- "Explain this change to me as if I'm not a developer, in three sentences." If it can't, the change is probably too big.
Make the AI keep your context for you
The biggest time-saver is not having to re-explain yourself. Coding agents such as Claude Code read skill files and project instructions automatically; ChatGPT, Claude and Gemini have Projects, custom GPTs and Gems that hold your background.
Our packs for builders — for example the AI vibe coder pack and the coding agents and MCP pack — come as ready files: prompts for each step of a build, and skills you can drop into a Claude Project or a .claude/skills folder. Each pack also includes a start file, 00-GIVE-THIS-TO-YOUR-AI.md: give it to your AI and it asks about your project once, remembers the answers, and fills them into every file in the pack. Browse all packs or start with the free prompts.
The short version
- 100,000+ new projects a day on Lovable alone; ~121M new GitHub repos and 36M new developers in 2025.
- Most AI-built projects stall for the same reasons companies' AI pilots do: no written goal, no definition of done, lost context, steps that are too big.
- A one-page brief — purpose, users, never-do list, decisions made, definition of done, decisions log — fixes most of it.
- Small, checkable steps beat one big ask every time.
Questions people ask
How many apps are built with AI every day?
There is no single total, but individual platforms publish their own figures. Lovable reported more than 100,000 new projects per day in 2025, and GitHub's Octoverse 2025 counted about 121 million new repositories that year — roughly 230 a minute.
Why do vibe-coded projects break so often?
Usually because the goal, the rules and the decisions already made exist only in the builder's head. The AI re-decides them each session, so changes undo earlier work. A short written brief with a definition of done for each step prevents most of it.
What should I put in a brief for an AI app builder?
One sentence of purpose, who uses it and what each user must be able to do, what it must never do, decisions already made (stack, payments, hosting), the definition of done for the current step, and a running log of decisions from earlier sessions.
Put this into practice
Grab a ready-made prompt from the Library, or a pack for your trade, fill in your details and take it to the AI you already use.
Keep reading
AI in 2026: 900 Million People Use It Every Week — So Why Do Most Results Disappoint?
ChatGPT has 900M weekly users and 88% of firms use AI, yet most see little return. What the 2025–26 research says, and how better prompts close the gap.
ChatGPT Prompts for Developers: Debug, Refactor & Review
35 copy-paste ChatGPT prompts for developers: debugging, code review, refactoring, tests, docs, and architecture. Paste your code, get reasoning.
RCTCO Prompt Framework: Role, Context, Task, Constraints, Output (+ Template)
RCTCO stands for Role, Context, Task, Constraints, Output — a 5-part prompt structure for ChatGPT, Claude and Gemini. Definition, worked rewrite and a copy-paste template.