How to Build an App With Replit: 2 Prompts, Tested Live

I had this app idea sitting in my notes for months. You paste in your resume, paste in the job post, pick a length, and get back a cover letter that actually sounds like a person wrote it. So I finally built it. Then I built it again, a completely different way, to test the one thing almost everyone gets wrong about AI app builders: how much detail your first prompt actually needs.

Both versions are real, hosted, working apps with logins and a database. Not demos. One came from a short prompt that made the agent interview me first. The other came from a single long spec I wrote out in advance. Same app, two ways to build an app with Replit, running side by side.

If you don't care about cover letters, stay anyway. The part that matters here is the prompting, and it transfers to any app builder you use. This is how you get the first version to come out close to the thing in your head instead of almost right and then broken.

Why Replit for this#

I built and hosted the whole thing on Replit. You describe what you want, and it writes the app, sets up the database, handles the logins, and puts it online with a real URL. That last part is the reason I keep coming back: the apps I host on Replit cost me about four cents last month, because I'm not burning compute when nobody's using them.

Replit sponsored this one, and I'm genuinely thankful for the early access, but I've been paying for and building on Replit for over a year now. I only partner with tools I already use.

Method 1: Question-First#

The first approach hands the agent a short, open-ended prompt and tells it to interview you before it writes a single line of code.

The prompt asks the agent to figure out who the app is for, the core features, what data needs to persist between sessions, whether users need accounts, and, most importantly, what you are NOT building so it doesn't go overboard. You paste it in, toggle on planning mode, and answer a handful of questions.

Three minutes of back and forth later, I had a full plan: the screens, the saved data, what the tool deliberately leaves out, and the key technical calls. It suggested Clerk for sign-in. I told it to use Replit's own auth instead, with email and Google login, just to see how easy that was to wire up. It was one sentence and done.

This is the move when your idea is fuzzy, or when you're new to Replit and don't know what you're even allowed to ask for. You approve the plan before any code exists, so you almost never end up with something wildly off.

Method 2: One-Shot PRD#

The second approach is the opposite. You write the entire product brief up front, one long spec, and let the agent build the whole thing in a single pass. No planning mode, because the prompt already spells everything out: the target user, the exact flow, the AI generation rules, the system prompt, what happens after generation, accounts and limits, guardrails, and the design.

I kicked this one off while the first agent was still working, so both builds ran at the same time. That alone is worth trying once, watching two agents build two versions of the same app in parallel.

The tradeoff is real: if your spec is wrong, you get a polished version of the wrong app. This is the approach for when you can already picture the screens and you know your non-negotiables, especially auth and spend limits.

What it cost, and where it broke#

Both builds landed close. The question-first version took about 12 minutes and $2.67 for the initial build. The one-shot version took 13 minutes and $3.69. Across the whole session, including the fixes below, the total was $13.51, roughly $6 to $7 per app, and all of it inside the Replit plan I already pay for. At that price the cost of experimenting is basically zero.

Then I tested them on a real LinkedIn job post, and one of them broke on camera.

I pasted in the resume and the job description, set the target length to 325 words, and hit generate. It gave me 25 words. Not 325. A hard cutoff.

This is the moment most people start arguing with the agent across five messages and dig a deeper hole. I sent one prompt instead: I told it exactly what I tested, that the generation cut off after 25 words instead of 325, and asked it to investigate and fix it so it builds the full, well-formatted letter tailored to the job and resume. One follow-up. It found the error, and the next generation came back as a complete, clean letter.

The other version had a smaller issue: the first PDF export looked squished, all one font size. Same fix pattern. One specific prompt describing what was wrong, and it rebuilt the formatting for both the editor and the download.

That's the honest friction point with any of these builders. The first pass is rarely production-polished. What separates people who ship from people who rage-quit is knowing that a build is a draft, and that one precise follow-up beats five vague ones. When it really won't cooperate, you roll back to a checkpoint instead of talking it in circles.

Which one wins#

Here's the surprise. After both fixes, the two apps were close to indistinguishable in quality. Same clean layout, same logins, same saved history, same genuinely good letters with varied sentence length and no giveaway AI tells.

So the answer to "how much detail does your first prompt need" isn't "as much as possible." It's this: the number of prompts barely mattered. The quality of the thinking did. A vague one-shot spec would have lost badly. A good one matched a full back-and-forth interview.

If you're new to this, start question-first on your first project. The interview teaches you exactly what the agent needs to know, which is the same knowledge that makes your one-shot specs good later. Once you can picture the screens, write the spec and let it rip.

Turning it into a business#

Because both apps ship with real logins and a database, the jump to a paid product is short. You connect Stripe, add tiers, and gate usage, say 10 free cover letters a month, then a few dollars a month for more. That protects you from a stranger running up your API bill and turns the tool into passive income. In the video I do it with a single prompt: add Stripe and a pro tier with 100 generations a month. If you want the full path from a free MVP to a real micro-SaaS, I break that down in building a micro-SaaS with Replit's AI agent.

I published one of the two versions so you can actually use it. Try the finished AI cover letter writer, and tell me which build you'd have put your name on.

Get both prompts#

The two prompts I used, the full question-first starter and the complete one-shot PRD, are in a free playbook. It also covers how Replit's agent actually works, the eight rules that get you better output, and the grounding block that stops the AI inventing job titles you never had. Copy them, swap your details, point them at whatever app is sitting in your notes.

If you want to see the same agent build three different apps in one afternoon, I did exactly that with Replit Agent 3, including an AI automation that emails me every Monday.

Watch the full build on YouTube: https://youtu.be/Wk-x0TNEvCw

This post is sponsored by Replit and contains affiliate links. I only recommend tools I actually use.