In my first vibe coding project ever, I built Fallacy Detector, a Chrome extension that uses Claude Sonnet 4 to flag cognitive biases and logical fallacies in news articles.
In just two weeks, I went from dozens of app ideas to a working product with:
15 detectable biases
Multi-run averaging for consistent results
Fast (6 s) and Accurate (17 s) modes
Clean, intuitive UI
90% reduction in score variance
Download a local version of the extension yourself (you'll need a Claude API key) here: GitHub
Once you've tried the app, provide some feedback: Feedback
I started October not as a developer, but as a product manager hoping to learn by building. My goal: create an AI-powered app from scratch—something real, not just an idea on a whiteboard.
To get there, I used what I call my “island of lost startup ideas”. Inspired by Greg Isenberg’s method, I brainstormed 37 app ideas with Claude and ran them through a Reddit Analysis Super Prompt to see what might actually resonate.
After some AI-assisted analysis and reflection, three contenders emerged:
An AI gift generator
A curated non-alcoholic beverage experience
A cognitive bias Chrome extension
The last one won out. It felt meaningful, buildable, and relevant to a world drowning in misinformation.
Bias in news isn’t new—but what struck me is how subtle it often is. It’s not about fact versus fiction anymore, but framing, tone, omission.
So the question was:
Could AI help highlight manipulation—not tell us what to think, but when something’s nudging how we think?
That’s the idea behind Fallacy Detector.
This was my first time vibe coding anything substantial. I chose Claude Sonnet 4 because it felt the most intuitive AI collaborator for a non-developer.
Even so, there were bumps. Sometimes Claude said it had included a particular UI element or function—but it wasn’t actually in the code. Debugging those invisible “ghost features” became part of my daily learning curve.
The MVP was simple:
Chrome extension
Context menu to analyze selected text
API call to Claude
Basic popup for results
It worked—but inconsistently. The same paragraph returned bias scores of 7, 4, and 6. Variance = 3 points ⚠️
Large models can give different outputs from the same input. So I implemented multi-run averaging—running the analysis three times and averaging results.
Variance dropped to zero.
Consistency achieved. Speed sacrificed. (~15 s total per analysis)
Fifteen seconds feels slow when you’re browsing. So I created two modes:
Fast Mode (6 s): 1 analysis, cheap and quick
Accurate Mode (17 s): 3 analyses, rock-solid stable
A hybrid design let users start fast and “upgrade” accuracy on demand.
As a first-time builder, my prompt initially crammed all settings behind one “Save” button. Confusing.
I redesigned the interface into clear cards:
Save API Key
Update Settings
Status Overview
It instantly felt more professional—and trustworthy.
Before:
After:
Every product needs personality. Fallacy Detector’s icon became a retro pixel‑art brain floating in space—symbolizing intelligence, curiosity, and play.
As the only user, I tested the extension across 40 news sites:
CNN, BBC, InfoWars, The Guardian, NYT, and local outlets.
The results were telling:
Extreme bias? Easy to catch.
Moderate bias? Needed averaging.
Speed? Always preferred Fast mode.
Prompt refinement became critical—especially distinguishing article bias from quoted source bias, which early versions misinterpreted.
Illustration suggestion: screenshot of popup interface showing article score and biases detected.
Stack:
Chrome Extension (Manifest V3)
Claude Sonnet 4 API
Chrome Sync Storage
Vanilla JS + CSS
Architecture:
Background: API calls + averaging logic
Content Script: popup + page interactions
Popup UI: settings and bias selection
Temperature 0 reduces randomness – variance ± 0.5 points. (In AI models, temperature is a parameter that controls the randomness of predictions. A temperature of 0 means the model will always pick the most probable output, reducing variance and making the system more deterministic. However, this comes with a trade-off of ± 0.5 points in variance, which could mean a slight loss in the model’s ability to explore different possibilities.)
Three‑run averaging = sweet spot for accuracy.
Speed > Perfection – fast feedback beats flawless precision. Nothing is more annoying than waiting an elongated time period for a result.
Prompt clarity is everything. It took me 20 prompts because…they needed clarity.
Good design, even in a v0, matters. Improving the design made me want to use the tool.
Claude sometimes omitted code lines it claimed to include. I learned to double‑check everything.
As a solo tester, I had to guess what users might find confusing.
Building from scratch taught me more about debugging mindset than syntax.
90% reduction in score variance
Fast vs Accurate: 6 s vs 17 s
Cost per analysis: $0.006 – $0.018
15 biases implemented
Working prototype tested on 40 sites
I started this journey wanting to build “an AI app someday.” Two weeks later, I had one.
Fallacy Detector isn’t perfect—but it works. More importantly, it represents my leap from ideation to creation.
If there’s one takeaway from my first project, it’s this:
It’s fairly easy to get started now. But it is not extremely easy - there is quite a bit of friction to overcome and I don’t anticipate it being so easy that everyone builds an app. I also think that building is of limited value if you don’t have a confirmed hypothesis - I didn’t gather any direct user research prior to building an mvp. But I did try to build for myself, and know there is a desire to more easily understand how news is used to manipulate.
Download a local version of the extension yourself (you'll need a Claude API key) here: GitHub
Once you've tried the app, provide some feedback: Feedback
Share Dialog
nftmatrix
Support dialog
No comments yet