Ideas
The honest problem with "here are 25 things to build with AI" lists is that half the ideas are lies. They look weekend-sized and then you hit auth, or billing, or realtime sync, and your Saturday becomes a fortnight. We've shipped enough with agents to know which ideas actually finish, so this list is ranked and labeled by what agents ship well today — not by what sounds impressive.
Read the label on each. "Weekend" means a coding agent and a focused person can genuinely land it in a day or two. "Looks easy, isn't" means there's a hidden hard part — usually auth, payments, or realtime — that eats the timeline. Build the honest ones first.
The tiers
| Tier | What it means | Realistic time |
|---|---|---|
| Weekend | Agent ships it well; no hidden hard part | 1–2 days |
| Long weekend | One real complication, otherwise clean | 3–5 days |
| Looks easy, isn't | Auth + billing + realtime hides inside | 2+ weeks |
Genuinely a weekend
- A single-purpose CLI utility. A tool that renames files by a pattern, converts between formats, or bulk-edits config. Agents are superb at CLIs — clear input, clear output, easy to test. Start here if you're new.
- An API wrapper library. A clean SDK around a public API that ships without a nice client. Well-scoped, well-known shape, easy to verify against the docs.
- A landing page with a waitlist. Static page, email capture to a table or a form service. A day, and the copy is the hard part, not the code.
- A Chrome extension that does one thing. Highlight-and-save, strip tracking params, reformat a page. The manifest and permissions model are the only friction, and agents know them.
- A scraper for one site. Pull structured data from a single source into a CSV or SQLite table. Scoped to one site, this is a clean afternoon.
- A static dashboard over a CSV or a read-only DB. Charts and tables, no writes, no auth. Genuinely fast because there's no state to defend.
- A Discord or Slack bot with a few commands. Slash commands, a couple of API calls, a reply. The platform SDKs are well-trodden ground for agents.
- A markdown-to-something converter. Notes to a styled PDF, a changelog to release notes, a spec to a checklist. Pure transformation, easy to test.
- A cron-driven digest. A script that runs daily, hits a few sources, and emails or posts a summary. The logic is small and the failure modes are visible.
- A personal API over your own data. A tiny read API over a database you control, no public auth. Scaffolds in an hour.
- A code-mod script. A one-off transform across a codebase — migrate an import, rewrite a deprecated call. Agents write these fast and you verify by running the build.
- A webhook relay. Receive a webhook, transform it, forward it somewhere. Small, testable, and useful more often than it sounds.
A long weekend — one real complication
- An internal tool with a login. A CRUD app for your team. Weekend-plus, because "with a login" is where hidden time lives even when the auth is simple.
- A data pipeline with a few stages. Ingest, clean, store, expose. Each stage is easy; wiring them so failures are visible is the real work.
- A scraper across several sites. Now you're handling different page shapes, rate limits, and retries. Still doable, but the edge cases multiply.
- A dashboard that also writes. Add "edit" to a read-only dashboard and you've added validation, optimistic UI, and conflict handling. Not hard, but real.
- A bot that remembers state. A Discord bot with a database behind it — polls, scores, queues. The persistence layer is the complication.
- A file-processing service. Upload a file, do something to it, hand it back. Upload handling and cleanup are the parts that bite.
- A public read API with keys. Rate limiting and API-key issuance turn a weekend read API into a long-weekend one. Manageable, but it's more than it looks.
Looks easy, isn't — respect these
- A SaaS with subscriptions. The moment you add auth and billing and a dashboard, you're in multi-week territory. Stripe webhooks alone will humble you. We wrote a whole map for this — see building a SaaS with AI agents before you start.
- A realtime collaborative anything. Multiplayer cursors, live docs, shared state. Conflict resolution and reconnection logic are genuinely hard problems that look like a checkbox on a feature list.
- A marketplace with two sides. Buyers, sellers, payouts, disputes. Each side is an app; the money movement between them is a third. This is a business, not a weekend.
- A chat app with presence. "Who's online," typing indicators, delivery receipts, offline queueing. The happy path demos in an hour and the real path takes a month.
- Anything with roles and permissions. "Admins can, members can't" spreads into every endpoint and every view. Permission logic is where security bugs breed; budget for it.
- A native mobile app with sync. Offline-first sync is one of the genuinely hard problems in software. Agents write the screens fast and then the sync layer eats the calendar.
How to actually pick one
Three rules keep you shipping instead of stalling:
- Start one tier below your confidence. Whatever you think you can build, build the version one notch simpler first. You'll finish it, and finishing is a skill you're training.
- Pick a boring stack the models know cold. Laravel, Next.js, SQLite or Postgres. Agents are dramatically better on frameworks they've seen a million times than on the clever thing you read about last week. Exotic stacks turn every task into a fight.
- Define "done" as something you can run. Not "it works" — "I ran
npm run dev, clicked the thing, and it did the thing." Verifiability is what keeps a weekend project from silently becoming a mess.
The move
Pick a genuine weekend item from the top of this list, scope it to the smallest version that's still useful, and ship it end to end before you touch anything from the bottom third. The "looks easy" tier isn't off-limits forever — it's off-limits until you've shipped two or three of the honest ones and learned where agents shine and where they need a hand. When you're ready for that first real end-to-end run, ship your first product with AI agents walks the whole path. Running a couple of agents in parallel to knock these out faster is what The Vibe Father is for, but every idea here ships fine with a single agent and a bit of nerve.