Red Berry, Dead Berry

Inspiration

The tech ecosystem is suffocated by toxic positivity. Builders are conditioned to optimize for success, ignoring the statistical reality that most projects fail. Learning from success is slow; learning from catastrophic failure is immediate. I approached this from a first-principles perspective, modeling the probability of project survival. The expected success of any complex system is the product of avoiding terminal failures:

$$P(\text{Success}) = \prod_{i=1}^{n} (1 - P(F_i))$$

Where $P(F_i)$ is the probability of a specific fatal error. Instead of building another dashboard to chase the $P(\text{Success})$, I built an engine to aggressively identify and isolate $F_i$. If you want to survive, you must look at the poisoned berries others have eaten and avoid them.

What it does

Red Berry, Dead Berry is an inversion of standard productivity tools. It is a brutalist, single-page web application that acts as a failure prediction engine.

1. The Morgue: You input a goal or project idea.
2. The Autopsy: The AI instantly outputs the three most statistically probable ways your specific idea will die, stripping away all encouragement and fluff.
3. The Warning Signs: These failure vectors are saved to a public database. The community validates these predictions by clicking "I Died From This," crowdsourcing the most lethal traps in any given domain.

How I built it

I utilized a vibe coding methodology, leveraging MeDo to translate my architectural constraints directly into full-stack logic.

• Architecture: I designed the backend logic through MeDo, enforcing strict relational schemas between the Graveyard_Queries and Autopsy_Reports tables.
• AI Logic: I engineered the LLM integration to bypass standard query matching by forcing the AI to append a Category_Tag to every submission. This allowed me to cluster similar failures in the database without complex vector search.
• UI/UX: I demanded a strict, hostile design system from the prompt engine—utilizing only deep crimson, charcoal, and white.

Challenges I ran into

The primary friction point was the LLM's fundamental training. Modern AI models are heavily weighted toward "helpfulness" via RLHF (Reinforcement Learning from Human Feedback). When asked to predict failure, the initial MeDo generations kept trying to append "But don't give up!" or offering unsolicited pivot strategies. Forcing the LLM into a strict, analytical, and ruthless posture required aggressive system prompting and multiple iterations to reliably break the positivity bias.

Additionally, forcing a natural-language builder to execute real-time state changes—like the asynchronous upvote counter—without triggering full-page DOM reloads required highly specific, atomic prompting.

Accomplishments I'm proud of

I shipped a fully functional, stateful application without writing boilerplate. I successfully inverted the standard hackathon paradigm of building "helpful" tools by proving that negative-space analysis holds massive utility. The brutalist UI perfectly matches the application's core logic, creating a cohesive, zero-friction user experience.

What I learned

Vibe coding is not a substitute for architectural discipline. Natural language is simply a new compilation step. If your internal logic is flawed, the AI will generate flawed code at unprecedented speed. I learned that prompting for constraint (telling an LLM what not to do) requires far more precision than prompting for generation.

What's next for Red Berry, Dead Berry

The immediate next step is engineering the viral loop. I will integrate a "Warn the Others" feature, allowing users to automatically publish their autopsy reports to social feeds to drive engagement. Following that, I plan to aggregate the Autopsy_Reports data to output macro-level failure trend analysis across different sectors (e.g., identifying the leading causes of death in consumer SaaS versus hardware startups).

Built With

• api
• css
• database
• html
• javascript
• llm
• medo
• prompt