One more Thing in AI
Posts
Painful lessons of a newbie vibe coder!

Painful lessons of a newbie vibe coder!

And how to be better at Vibe coding.

Renjit Philip
November 01, 2025 • Estimated Reading Time: 24 minutes

Date: 1-Nov-2025

Hey AI enthusiast,

I have been diving into the world of vibe coding and I thought of cataloging my experience for the benefit of others. My coding skills are 25 years old (from early 2000s).

Even though I did learn “modern” coding for my startup in 2018, I don’t consider myself anything more than a beginner.

So when the phenomenon of AI powered coding came up, I had to try it!

TL;DR

I spent 30 days building a financial planning tool for founders by “vibe coding”. Made every mistake possible. Burned through $127 in Replit credits. Got calculations wrong by 20%. Watched 23 "interested" founders become 2 signups.

But I also learned that:

(1) AI struggles with precision calculations, (2) founders don't want another spreadsheet—they want advice, and (3) sometimes less is more.

One founder offered $50/month for the simplified version. That's validation enough to keep going.

This was a great learning experience for me as I go to be on the other side of the table so to say. I spend time working with founders at accelerators and advising them about getting from zero to 1.

This was a good opportunity for me to “eat my own dog food”! Want to test out what I built? Here is the link» https://runwayai.futureu.co

Here's the full story and what I'd do differently.

In this edition:

Building a Financial Planning Tool for Founders: 3 …

Best,

Renjit

Building a Financial Planning Tool for Founders: 30 Days of Vibe Coding

A brutally honest journey from idea to MVP, including every mistake, pivot, and lesson learned along the way.

Reading time: 8 minutes
Best for: Founders building with AI, anyone considering vibe coding for serious products

The Problem That Obsessed Me

The VC leaned forward: "What if your churn drops by 2%?"

I watched the founder's face fall. His answer was buried somewhere in a 47-tab Excel model. It would take him 3 hours to calculate. The meeting momentum died right there.

I've seen this scene play out dozens of times. Founders spending hours, sometimes days, updating financial models. Breaking formulas. Forgetting which cells to change. Creating circular references that crash Excel.

The pain point I kept hearing:

"I spent 2 days building a financial model for investors. Then they asked 'what if churn drops by 1%?' and I had to rebuild everything."

The math is brutal: A large % of startups still use spreadsheets for financial planning and forecasting. Most founders hate every minute of it.

So I decided to build something better for the founders I mentor at accelerators like Founder Institute, Flat6Labs, Ignite Program, NSRCEL and so on.

Week 1: The Honeymoon Phase (Days 1-7)

Day 1: The Dream Begins

I dove into vibe coding with @Replit, convinced this would be straightforward. The AI influencers made it look easy on TikTok, right?

My initial vision:

AI-assisted financial cockpit ✨
Real-time data sync with QuickBooks, Stripe 📊
Scenario planning built-in 🎯
Investor-ready exports in seconds 📄

Estimated timeline: 2-3 weeks to MVP

[Narrator voice: It was not 2-3 weeks]

Days 2-7: Early Mistakes

Reality check came fast. I made three critical early mistakes:

Mistake #1: Assumed the AI agent operates in parallel

I gave multiple instructions by Replit Agent while it was working on my previous prompt:

"Make the dashboard cleaner"
"Add a night mode"
"Fix the calculation bug"

The agent stopped, absorbed all the prompts, got confused, and created a Frankenstein version that did none of them well.

Cost: 6 rollbacks, 3 hours lost, $23 in credits

Mistake #2: Underestimated UI complexity

Me: "Add night mode"
AI: Proceeds to make 47 changes
Result: White text on white background, buttons invisible, complete disaster

Patiently working through font/background mismatches took 3 days longer than expected.

Mistake #3: Wasn't precise with instructions

Bad prompt: "Make it more intuitive"
What happened: AI changed the entire navigation structure

Better prompt: "Change the 'Calculate' button color to blue (#0066CC) and increase font size to 16px"
What happened: Exactly what I wanted

The Magic Incantation

I discovered the key phrase for vibe coding for use with all vibe coding tools:

"Don't make any changes without confirming your understanding with me."

This simple instruction could have saved me $50+ in wasted iterations.

Stop Drowning In AI Information Overload

Your inbox is flooded with newsletters. Your feed is chaos. Somewhere in that noise are the insights that could transform your work—but who has time to find them?

The Deep View solves this. We read everything, analyze what matters, and deliver only the intelligence you need. No duplicate stories, no filler content, no wasted time. Just the essential AI developments that impact your industry, explained clearly and concisely.

Replace hours of scattered reading with five focused minutes. While others scramble to keep up, you'll stay ahead of developments that matter. 600,000+ professionals at top companies have already made this switch.

Join them today, for free.

The Gold standard for AI news

AI will eliminate 300 million jobs in the next 5 years.

Yours doesn't have to be one of them.

Here's how to future-proof your career:

Join the Superhuman AI newsletter - read by 1M+ professionals
Learn AI skills in 3 mins a day
Become the AI expert on your team

Start learning AI now

Week 2: Reality Check (Days 8-14)

Day 8-14: Travel and Trouble

Summer travel (to Japan), slowed me down, but also gave me perspective on what wasn't working.

Things I learned while vibe coding from airport lounges:

Hotel WiFi + Replit = frustration
TypeScript errors are harder to debug on mobile
The rollback button becomes your best friend

I would recommend Replit as your starting point for Vibe coding if you do not know much about coding and deploying applications. Replit handles all that easily.

Also it takes care of protecting your API keys and SECRETS which means no data leaks and high cloud or AI bills!

Plus it has a dedicated AI agent that you can chat with, understand code, review code and plan your work BEFORE you build. All of which I found quite useful.

Here is my affiliate link to test out Replit: https://replit.com/refer/renjitphilip

Day 15: Existential Dread

Tuesday, 3:47 PM: Recurring error to render the screen
Tuesday, 4:15 PM: Still debugging
Tuesday, 6:30 PM: Finally understand it's a TypeScript type issue
Tuesday, 11:45 PM: Still broken

Took a whole day just to understand the code.

Should I have told the agent to use Python (a language I somewhat understand).

Too late. The AI Agent was coding away in TypeScript. A language I don't really know.

Thank god for @Replit's rollback options.

Day 16: The Credit Crisis

Lesson learned: Replit credits run out FAST.

My spending pattern:

Week 1: $34 in credits
Week 2: $93 in credits (yikes!)

I was iterating rapidly, making changes, rolling back, trying again. Each iteration can cost $2-5 depending on complexity.

New rule: Budget $40/week maximum. If I hit the limit, stop and reflect on why I'm burning through credits. This is where the limit feature in Replit comes in handy.

The Stack Choices

I integrated:

@Brevo for system emails (free tier = win). Basically to send emails to users.
@MistralAI for the built-in AI advisor (seemed cheap at first...). I assumed that Open AI API calls would be too expensive.

Both seemed reasonable. The MistralAI choice would come back to haunt me.

Week 3: User Feedback Changes Everything (Days 15-21)

Day 17: Time for Feedback

After clearing up errors while traveling, I started hunting for testers.

Posted in 3 founder Slack channels:

"Building a financial planning tool that doesn't suck. Need critical feedback. DM if interested."

Response: crickets 🦗

Day 18: The Humbling

Finally got 1 friend and 2 founders to try the MVP.

The feedback was... brutal:

Issue #1: Financial calculations were off by 20%

A founder's CAC calculation showed $47 when it should have been $58.75.

That 20% error could have cost them their Series A round.

Why it happened: I had MistralAI calculate it based on vague instructions. The AI made assumptions about what "customer acquisition cost" meant.

Issue #2: Export feature crashed on larger models

Any model with >50 rows of data caused a memory overflow.

Issue #3: "Too many clicks to get to runway calculation"

The thing founders wanted most (runway) was buried 3 screens deep.

This was definitely harder than the AI influencers make it seem!

Day 20: Six Hours of Hell

The bug: LTV/CAC calculations were consistently wrong.

Time spent: 6 hours straight debugging

The cause: @MistralAI was interpreting "monthly churn" as "annual churn" in some scenarios, and "annual churn" as "monthly churn" in others.

When I asked it to "calculate customer lifetime value," it made assumptions:

Sometimes it used monthly churn rate
Sometimes it used annual churn rate
Sometimes it invented its own churn calculation

The fix: Be surgically precise with financial formulas.

Bad prompt:

Calculate LTV

Good prompt:

Calculate LTV as: (Average Revenue Per User * Gross Margin) / Monthly Churn Rate
where:
- Average Revenue Per User = Total MRR / Active Customers
- Gross Margin = (Revenue - COGS) / Revenue
- Monthly Churn Rate = Churned Customers This Month / Active Customers Start of Month

Critical Lesson Learned

Always validate AI outputs with manual calculations first.

I started keeping a spreadsheet (yes, ironic) to validate every calculation the AI generated.

Day 21: The Lightbulb Moment

One founder cut through all the complexity:

"I don't want another financial model builder. I just want to ask 'how do I extend runway by 3 months?' and get an answer."

So simple. So obvious.

I started pivoting immediately. Natural language queries → instant financial insights.

This changed everything. This is where I added a detailed chat functionality for the founders to ask questions about their financial model based on what they had entered so far.

Week 4: The Pivot (Days 22-30)

Day 22: Language Regrets

Regret of the month: Starting with TypeScript instead of Python.

When financial formulas get complex, I spent more time debugging syntax than building features.

Example that took 2 hours to fix:

// What I wanted:
const runway = currentCash / monthlyBurn;

// What TypeScript gave me:
Type 'number | undefined' is not assignable to type 'number'

Note to future self: Pick programming languages you actually understand for serious projects!

Day 23: The "Intuitive" Disaster

Replit credits burned through again. $11 in one afternoon.

I built a "scenario comparison" feature that took 47 iterations because I kept saying "make it more intuitive" without being specific.

Iteration 1: "Make it more intuitive"
AI changes the entire layout

Iteration 2: "No, not like that, more intuitive"
AI changes the color scheme

Iteration 3: "JUST MAKE IT INTUITIVE"
AI adds tooltips everywhere

...45 iterations later...

New rule: Define "intuitive" BEFORE asking for changes.

Better approach:

"Intuitive means: (1) Show both scenarios side-by-side, (2) Highlight differences in yellow, (3) Add a 'Switch to This Scenario' button on the right"

Day 24: The Harsh Reality of User Acquisition

Posted on 5 founder Slack groups asking for testers:

Results:

23 "looks interesting!" 😊
2 actual signups 😐
0 completed onboarding 😭

Conversion funnel:

Interest: 23
Signup: 2 (8.7%)
Onboarding completion: 0 (0%)

Making stuff people want ≠ Making stuff people actually use.

Day 25: The API Awakening

Founders kept asking: "Can it sync with QuickBooks?"

Me: "Sure, how hard can API integration be?"

[Researches QuickBooks API for 3 hours]

What's required:

OAuth 2.0 flow setup
Webhook validation
Data mapping between QuickBooks and my schema
Error handling for rate limits
Token refresh logic
Accounting rules validation

Realization: This isn't happening in vibe coding. Not anytime soon.

Day 26: Precision in Prompts

@MistralAI is great, but I learned you need VERY specific prompts for financial calculations.

Bad prompt:

Calculate runway

What AI does: Makes assumptions, might use wrong numbers

Good prompt:

Calculate runway as: current_cash / monthly_burn_rate
where monthly_burn_rate = (total_expenses - revenue) / months

Use these exact values:
- current_cash: $250,000
- total_expenses (last 3 months): $180,000
- revenue (last 3 months): $45,000
- months: 3

Show your work step by step.

What AI does: Gets it right every time

Precision matters in finance. No room for ambiguity.

Day 27: Rollback Appreciation Day

Used @Replit's rollback feature 12 times in one day.

What happened: I tried to add "smart defaults" for SaaS metrics (based on industry benchmarks).

Instead, I broke:

The calculation engine
The export functionality
The scenario comparison
User authentication (somehow?)

Time to fix manually: Unknown, gave up after 1 hour
Time to rollback: 30 seconds

Sometimes the best code is the code you don't write.

Day 28: The Hard Truth

After 4 weeks, I admitted what I'd been avoiding:

The full vision is too complex for vibe coding.

I scaled back to core MVP:

✅ Manual financial model builder
✅ AI advisor for benchmarks
✅ Basic scenario planning (3 scenarios max)
❌ Real-time integrations (later)
❌ Advanced version control (later)
❌ Multi-user collaboration (later)

Sometimes less is more.

Day 29: The Best Feedback

One founder's comment changed my entire approach:

"I don't need another spreadsheet. I need someone to tell me if my numbers make sense."

Tool → Advisor.

This was the pivot. Not another financial model builder. An AI-powered financial advisor that:

Validates assumptions
Flags unrealistic numbers
Suggests improvements
Answers "what if" questions

Sometimes users tell you exactly what to build if you listen carefully enough.

Day 30: One Founder Who'd Pay

Posted the scaled-back version. Got brutal but helpful feedback.

Then one founder said:

"This is the first time I've understood my unit economics without a finance degree. I'd pay $50/month for this."

That's validation enough to keep going.

The 30-Day Scorecard (Being Brutally Honest)

Month 1 Statistics

Development:

Days coding: 30
Replit credits spent: $127
Lines of code written: ~3,500 (mostly by AI)
Rollbacks executed: 73
Languages learned the hard way: 1 (TypeScript)

User Metrics:

Interested parties: 23
Actual signups: 12
Completed onboarding: 3
Completed models: 3
Would pay: 1
Revenue: $0

Key Wins:

🎯 Found product-market fit signal (1 founder willing to pay)
🎯 Validated core problem (founders hate Excel)
🎯 Discovered simpler solution (advisor > tool)

Key Losses:

❌ Built too much complexity first
❌ Didn't validate calculations carefully enough
❌ Chose wrong programming language
❌ Wasted $50+ on vague prompts

What I Learned About Vibe Coding

The Good ✅

Rapid prototyping: Went from idea to testable MVP in 2 weeks
Low initial investment: $127 total vs. $20K+ for developers
Quick iterations: Try, fail, rollback, try again in minutes
AI assistance: Great for boilerplate and standard patterns
No hiring headaches: No interviews, no management, no equity

The Challenges ⚠️

Credit management: Can get expensive fast with too many iterations
Complex calculations: AI struggles with precision in financial formulas
Integration limitations: Enterprise APIs need proper development
Language constraints: Stick with what you know or pay the learning tax
Precision required: Vague instructions = wasted time and credits
Debugging difficulty: Understanding AI-generated code is hard

Key Principles I Discovered

1. Be Surgically Precise with Instructions

Don't say: "Make it better"
Do say: "Change button color to #0066CC, increase font size to 16px, add 8px padding"

2. One Change at a Time

Never give the AI multiple tasks in parallel. Wait for completion, review, then proceed.

Bad workflow:

"Add dark mode"
"Fix the bug"
"Improve performance"

Good workflow:

"Add dark mode"
[Wait for completion]
[Review changes]
"Now fix the calculation bug in line 47"

3. Validate Everything

Especially financial calculations. Don't trust the AI. Check manually.

4. Use Rollback Liberally

It's there for a reason. I rolled back 73 times in 30 days. No shame.

5. Listen to Users

They'll tell you what to build. The founder who said "I need someone to tell me if my numbers make sense" gave me the entire product direction.

6. Know When to Pivot

Not everything can be vibe coded. Enterprise integrations, complex calculations, and performance-critical features need traditional development.

If I Started Over Tomorrow

What I'd Keep ✅

Vibe coding for rapid prototyping: Still the fastest way to test ideas
Brutal honesty in user testing: Critical feedback > polite feedback
The pivot to "advisor not tool": This is the winning insight
Daily build-in-public updates: Accountability + audience building

What I'd Change 🔄

Start with Python, not TypeScript
- Language I understand
- Better for data/calculations
- Easier to debug
Get 10 user interviews BEFORE building
- Would have discovered the "advisor not tool" insight on Day 1
- Could have skipped 2 weeks of wrong direction.
- Funnily enough I take a class on “Customer Validation” at an accelerator. But did not follow my own advice!
Set a 2-week deadline for MVP v1
- Prevents feature creep
- Forces prioritization
- Builds shipping muscle
Budget $200 for Replit, hard stop
- If I exceed budget, something's wrong
- Forces better prompt engineering
- Prevents death by 1000 iterations
Build manual process first, automate second
- Would have discovered QuickBooks integration isn't MVP
- Could have launched 2 weeks earlier

What I'd Skip Entirely ⛔

Night mode - Nobody asked for it, took 3 days
Perfect UI - Founders want function over form
Integration promises - Validate manual flow first
Advanced features - Get 10 paying users before adding more

The One Thing I'd Do Differently

Talk to 50 founders BEFORE writing a single line of code.

Not 5. Not 10. Fifty.

Ask them:

Show me your current financial model
What takes the longest to update?
What question do investors always ask?
What would you pay for a solution?

This would have saved me 2 weeks and $$ in wasted effort.

What's Next?

The Smart Approach: Progressive MVP

I'm not building everything at once anymore. Here's the realistic path:

Phase 1: Validate with Vibe Coding (Current - Weeks 5-8)

Build:

✅ Manual financial model builder
✅ AI advisor for validating assumptions
✅ Basic scenario planning (3 scenarios)
✅ Export to Excel/PDF
✅ Natural language Q&A

Goal: Get to 10 paying customers at $50/month

What I'm trying to prove:

Core value proposition works
Users prefer this vs. spreadsheets
AI advisor adds real value
Basic unit economics engine is sufficient

Success criteria:

10 paid subscribers
<10% monthly churn
NPS >40
At least 5 would "very disappointed" if product went away

Phase 2: If It Works, Go Pro (Weeks 9-24)

After 50-100 paying founders, consider:

Hiring fintech-experienced developers
Building real-time integrations
Adding enterprise security (SOC 2)
Scaling infrastructure

Budget: $50K-100K (bootstrap or small angel round)

When to Graduate Beyond Vibe Coding

Vibe coding is great for:

✅ Rapid prototyping
✅ Basic CRUD operations
✅ AI integration (OpenAI/Mistral API calls)
✅ Export functionality
✅ User feedback and fast iteration
✅ Landing pages and marketing sites

Where you'll hit walls:

❌ Complex financial formulas (cohort retention curves, NPV calculations)
❌ Enterprise API integrations (OAuth flows, webhook handling)
❌ Background jobs for data sync
❌ Multi-tenant security architecture
❌ Performance optimization (sub-300ms queries)
❌ Real-time collaboration features

The graduation threshold: When you have 10+ paying customers asking for features that vibe coding can't deliver.

If You're Building with Vibe Coding: A Checklist

Save yourself time and money. Use this:

Before You Start

[ ] Choose a language you actually understand
[ ] Set a credit budget ($X per week max)
[ ] Define "done" for your MVP (write it down)
[ ] Find 3 users who will test, not just say "sounds interesting"
[ ] Interview 10+ potential users first
[ ] Write down the #1 problem you're solving

While Building

[ ] One prompt at a time, wait for completion
[ ] Define vague words ("intuitive", "clean", "simple")
[ ] Validate ALL calculations manually
[ ] Use rollback guilt-free
[ ] Track credit spend daily
[ ] Screenshot working versions before major changes

When to Pivot

[ ] If the same error persists after 5 attempts
[ ] If you're explaining more than building
[ ] If test users can't complete core flow
[ ] If enterprise features keep coming up
[ ] If you've spent >$200 without paying users

Want to Help?

I'm looking for 10 founders to:

Test the MVP for 1 week
Jump on a 15-min feedback call
Get free lifetime access if you help shape v1

Requirements:

✅ Pre-seed to Series A stage
✅ Currently use Excel/Sheets for financial planning
✅ Willing to be brutally honest (critical feedback only)
✅ Can commit to testing for 1 week

What You Get:

🎁 Free lifetime access (worth $600/year)
🎁 Priority feature requests
🎁 Direct access to me via WhatsApp
🎁 Your logo on the "Built with Founders" page

Spots available: 7/10 (3 already filled) as of Nov 1, 2025

How to apply: DM me on Twitter @RenjitPhilip. Want to test out what I built? Here is the link» https://runwayai.futureu.co

"This is the first time I've understood my unit economics without a finance degree. I'd pay $50/month for this."
— Mike R., SaaS Founder, Pre-seed

"You answered my investor question in 30 seconds. It would have taken me 3 hours in Excel."
— Sarah K., Marketplace Founder, Seed

The Mission

Kill the 47-tab Excel financial model.

Every founder deserves:

✅ Real-time financial dashboards
✅ AI that explains the numbers
✅ Scenario planning in seconds
✅ Investor-ready exports instantly

The journey continues. Day 31 starts tomorrow.

Follow along: @RenjitPhilip #vibecoding #BuildInPublic

P.S. - If you found this helpful, share it with a founder friend who's drowning in Excel. They'll thank you later. Here is the link» https://runwayai.futureu.co

Want to reach a quality audience?

Reply

or to participate.