Last weekend I started vibe coding with Lovable and it was a blast!
Picture this:
- I'm at my computer prompting away, having fun learning how to use the Gemini API, GitHub, and Supabase
- Woo hoo! I'm just chit chatting with the AI agent, we're building (I think?)
- Then reality hits: 94/100 messages gone
I had a decision: should I pay to upgrade to Lovable's $50/month plan to get 250 messages? Hard pass.
I'm non-technical and cheap 💁♂️ I'd burn through that tier creating a mediocre MVP and face a $200 bill.
What’s a scrappy, low-budget "builder" to do?
TLDR;
- Assemble your AI "team" using free/cheap tools
- Use ChatGPT/Claude + a notes app to map the path
- FAFO across free tiers - pay only when you get it
- Embrace the deep pain of debugging
Step 1: Build an AI Team
Before you burn through your free-tier limits or overpay for a premium plan, assemble an AI team to progress with minimal spend.
Cofounder: Generalist Chat
- Tools: ChatGPT, Claude, Gross
- Cost: $20/mo
- Role: Lead Engineer, brainstorm partner, planning, and first-pass prompts
- Tip: Create a dedicated project
Eng. Team: CodeGen Applications
- Tools: Lovable, v0, Replit, Bolt, Devin, whatever, they're commodities for our purposes
- Cost: Free tier, then 1st paid tier is ~$20/mo.
- Role: Code generation, UI mockups, telling you things are broken
- Tips:
- Try them all, see what you like most
- Pick only one to eventually pay for
You're going to strategically use free tiers alongside one paid 'workhorse' general LLM chat to cover your needs.
Step 2: Create a clear description of your vision and MVP
Tool: ChatGPT/Claude
Before you prompt a single line of code, define exactly what you’re trying to build. As a wise person once said, "always begin with the end in mind."
If your vision is vague, codegen applications will waste your limited messages generating junk that doesn't fit what you want.
You should be able to tell ChatGPT/Claude the basics like:
- What problem you're solving
- Who it's for
- What the UI/UX should be like
- What the MVP includes
Then use the same tool to help you get clearer:
“Help me define a minimum viable product for [your idea], focused on core functionality only. Tell me what isn't clear yet.”
Step 3: Generate rough draft prompts
Tool: ChatGPT/Claude
Now that your vision is clear, it's time to use ChatGPT/Claude to generate the series of prompts to make your vision come to life.
You want individual, tightly-scoped prompts. See: Incremental prompting
Here is what I used and got ok results:
I would like you to tell me the best series of prompts that would help a tool like v0 or replit create that exact vision you just laid out.
I ended up with an initial 30 prompts that laid out a path forward.
Step 4: Edit the prompts
Tool: Notes app or textedit
Once you've generated your rough draft prompts, don't use them yet.
First, clean them up to ensure every word drives toward your MVP and nothing else.
This was my process
- Dump the prompts into a note-taking app or textedit.
- Manually review each prompt
- Delete lines or entire prompts if they bloat your MVP with unnecessary features
- Capture ideas for later.When you find cool but off-scope prompts move them to a Product Roadmap section for future builds
- Refocus vague prompts into specific, outcome-driven tasks.
Unedited prompts = wasted messages and bloated code.
Plus, editing prompts in a text file is cheaper and easier than attempting to edit code.
Step 5: Prepare your feedback loop
You're now about to test your prompts and learn a hard truth: most things you “build” will be very very broken.
To prepare, set up a feedback loop to evaluate, refine, and reprompt the codegen applications without burning messages needlessly.
You're going to spend most of your time debugging, so strap in.
Here's my approach today 👇
#1) Screenshot the app error and file structure
#2) Share both screenshots with ChatGPT/Claude, ask:
- What's the issue?
- What file should I review to diagnose the issue?
- What needs to change to resolve the issue?
- What other files or areas of the app might be involved?
- What should I prompt Lovable so that it solves this issue?
- Act as if you're speaking to a junior engineer
#6) Rinse and repeat
The goal is to offload most of the thinking to an unlimited-use tool and avoid wasting limited codegen application messages on trial and error.
That said, this process barely works, which shows just how far codegen applications have to go.
For persistent issues, I dig through API docs, scour forums, and ask for help. AI gets me partway, but manual effort is unavoidable.
Step 6: Test your prompts and feedback loop with free tiers
Tools: Your codegen eng. team
Now that you’ve got rough prompts and a pre-planned feedback loop, it’s time to take them for a spin across alllllll the free tools you're curious to try.
You'll hit usage limits and have to move on. That's fine, starting over often is part of the game.
So get comfortable with that!
Use free tools as your sandbox until you are ready and confident that spending is worth it.
Your goals are simple and fun:
- Learn about yourself and your vision
- Catch weak prompts before they cost you
- Learn how different tools interpret your prompts
- Learn how to debug issues in a no-pressure environment
As a bonus tactic, take detailed notes as you go along about which prompts worked, what broke, and how you fixed it.
Over time, you’ll build your own prompt library, a personal cheat sheet that saves hours later.
Cheap, Non-Technical, and "Building"
You don’t need to be technical or drop $200/month to start "building" with AI.
You need a clear vision and a process that lets you fail fast without paying up front.
Is it perfect? Oh god no.
But it’s good enough to start, and starting is everything!
.jpg)
