The Claude → Replit Hackathon Prompt: Turn a Problem Statement into a Working Prototype in Minutes

Source: Devshree's Substack · Author: Devshree Bharatia (Develop with Devshree) · Date: Apr 22, 2026 · Original post

---

The big idea

AI hackathons reward speed and polish. The bottleneck for most participants isn't writing code — it's translating a vague problem statement ("build something for mental health") into a concrete, judge-impressing prototype before the clock runs out.

Devshree's trick is a two-step pipeline that splits the work between two AI tools, each doing what it's best at:

1. Claude plays the role of a senior software architect. You hand it your hackathon problem statement, and it produces a detailed, structured app brief — the kind of spec a tech lead would write before kicking off a sprint.
2. Replit Agent plays the role of the implementing junior developer. You paste the brief in, and it scaffolds and builds the actual working app.

The mental model: don't ask one AI to both think and type. Use Claude to remove ambiguity, then use Replit Agent to execute against an unambiguous spec. The quality of what Replit builds is bounded by the quality of the brief it receives — so the leverage point is the brief, not the code.

The three-step workflow

1. Paste the prompt below into Claude, replacing the placeholder with your hackathon problem statement.
2. Copy the brief Claude generates and paste it into Replit Agent.
3. Wait a few minutes — you get a complete, runnable application.

That's it. The whole edge is in the prompt's structure.

The Claude prompt (full text)

You are an expert software architect who specializes in writing detailed app briefs for AI coding agents like Replit Agent.

I am participating in a hackathon. Here is my problem statement: [PASTE YOUR PROBLEM STATEMENT HERE]

Your job is to generate a single, detailed, structured brief that I will paste directly into Replit Agent to build a fully working prototype.

The brief must include:

1. APP OVERVIEW

• App name & one-line description
• Core problem it solves
• Target user

2. FULL FEATURE LIST

• List every feature the app should have, from core to nice-to-have
• Prioritize features that will impress hackathon judges

3. TECH STACK

• Specify the exact technologies, frameworks, libraries and tools Replit Agent should use (keep it beginner-friendly and fast to build)

4. PAGES & USER FLOW

• List every page/screen in the app
• Describe what each page does and how users navigate between them

5. UI & DESIGN INSTRUCTIONS

• Color scheme, fonts, layout style
• Any specific UI components or animations to include
• Must look polished and modern out of the box

6. DATA & APIS

• Define the data models/schema
• List any free APIs to integrate with exact endpoint details
• Include any dummy/mock data to pre-populate the app

7. STEP-BY-STEP BUILD INSTRUCTIONS FOR REPLIT AGENT

• Write clear, sequential instructions Replit Agent should follow to build the entire app from scratch
• Be extremely specific so there is no ambiguity

Output ONLY the Replit Agent brief. Nothing else. No explanations, no commentary. Just the ready-to-paste brief.

Why each section of the brief matters

The seven sections aren't arbitrary — they each close a specific gap that causes AI-generated apps to look half-finished:

• App overview forces the model to commit to a name, a one-liner, and a target user. This anchors every later decision (a productivity app for students looks very different from one for executives).
• Full feature list, judge-prioritized. Most generated apps fail at hackathons because they implement only the obvious feature. Telling Claude to prioritize what impresses judges nudges it toward the wow-factor feature (live AI integration, real-time collab, a creative twist) instead of yet another CRUD form.
• Exact tech stack, beginner-friendly. Replit Agent works best on conventional, well-documented stacks. Pinning the stack avoids the failure mode where the agent picks something exotic and gets stuck on dependency or deploy errors. "Fast to build" is the constraint that matters in a hackathon, not "best at scale."
• Pages & user flow. Without this, agents tend to build a single-page demo. Listing every screen and the navigation between them is what turns a demo into something that feels like a product.
• UI & design instructions. Judges form an opinion in the first ten seconds — usually from the visuals. Specifying colors, fonts, layout, and animations up front means Replit doesn't ship the default Tailwind-grey look that screams "AI generated."
• Data & APIs, with mock data. Two practical wins: free APIs (with exact endpoints) prevent the agent from inventing fake URLs, and pre-populated dummy data means the app looks alive the moment it loads. An empty app demoed to judges feels broken even when it works.
• Step-by-step build instructions. This is the section that turns the brief from a description into an executable plan. Replit Agent works much better with sequential, unambiguous steps than with a wishlist of requirements.

The closing line — "Output ONLY the Replit Agent brief. Nothing else." — is also load-bearing. Without it, Claude tends to wrap the output in conversational filler ("Here's your brief! Let me know if you'd like changes...") that pollutes what you paste into Replit.

How to actually use it

1. Have your hackathon problem statement ready, even if it's rough — Claude will fill in the gaps.
2. Paste the prompt into Claude, swap in your statement, run it.
3. Skim the brief once. Tweak anything obviously off (wrong target user, missing feature you care about) by replying to Claude rather than editing by hand — keep the structure intact.
4. Paste the final brief into Replit Agent and let it build.
5. Use the remaining hackathon time for the things AI still can't do for you: a clean demo script, a sharp pitch, and one or two custom touches that make the project feel personal.

Final thought

The takeaway isn't really "this exact prompt." It's the pattern: separate spec generation from implementation, and over-specify the spec. A vague prompt to Replit gives you a vague app. A Claude-generated brief that names the colors, lists the endpoints, and orders the build steps gives you something that looks like a real product — which is usually all it takes to win.

Partial summary note: the original post is largely the prompt itself plus framing, all of which is captured above.