Vibe-code an AI agent that posts to social media

What you'll learn

How to vibe-code a real product: you describe what you want in plain language, an AI coding tool writes the code, and you steer it until a working app is live on the internet. Here the product is an AI agent that writes social media posts and publishes them.

What vibe coding is, and is not

You are the product owner and the AI is the engineer. You do not need to write the code yourself, but you do need to know what you want, test what you get, and say clearly what is wrong. Tools that work this way include Claude Code, Cursor, Lovable, Replit and v0. Pick one and stay with it.

Build it in this order

Ask for one piece at a time, run it, and only then ask for the next. A prompt that asks for everything at once gives you something you cannot debug.

1. A page that works. "Build a web app with one page: a text box for a topic, a button, and a list of drafts." Run it. See it in the browser.
2. The writer. "When I press the button, call a language model and write a post about the topic, for LinkedIn, in a friendly professional voice, under 600 characters." Try five topics. Fix the voice by telling the AI what you did not like.
3. Approval. "Each draft has Approve, Edit and Discard. Nothing can be published until it is approved." This is the step that makes it safe to hand to a company.
4. A real account. "Connect my own account on one platform through its official API and publish an approved post." Do the platform's developer setup yourself; the AI tool can walk you through it.
5. A schedule. "Let me choose a time for an approved post, and publish it then."
6. A history. "Show every post, where it went, when, and a link to it. Never publish the same draft twice."
7. Deploy. "Deploy this so it has a public URL." Vercel, Netlify, Replit and Lovable can all host it for free.

How to steer when it goes wrong

• Paste the exact error message back to the AI. Do not describe it from memory.
• Say what you expected and what happened instead.
• If it goes in circles, go back to the last version that worked and ask for a smaller step.
• Commit to GitHub after every step that works. That is your undo button.

Keep it safe

• Publish only to accounts you own, or a test account made for this challenge.
• Keys and tokens go in environment variables. Tell the AI this in your first prompt, and check the repo before you share it.
• Approval is on by default. An agent that posts on its own the first time it runs is a bug.

Keep your prompts

Save the prompts that mattered and what you changed after each one. That build log is part of what you hand in: it shows how you think, which is what a company hiring you wants to see.

When someone else can open your URL, type a topic, approve the draft and see it appear on a real account, you are ready to submit.