# Promptimize

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Gabriel Danandjaja](https://cerebralvalley.ai/u/stormragemc)
- **GitHub:** https://github.com/stormragemc/Promptimize
- **Demo video:** https://youtu.be/I1rdD-wBmR0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2624

TRY PROMPTIMIZE: https://promptimize-steel.vercel.app/

NEWEST PATCH 

Updated main branch and VERCEL deployment (5/11/2026 9.50 AM SGT): Enabled Muse Image 1.0 to unconditionally (very minimal context needed) turn a user’s promptbase into a visual explanation of how its AI assistant could load context more efficiently. It can now clearly explain suggested prompt retrieval algorithms and no longer requires the minimum context condition which was supposedly shown as a prerequisite in the DEMO video.

PROJECT SUMMARY 

Meta’s Thin Orchestrator tech blog earlier this year described an AI agent whose mixed instructions caused it to apply rules from the wrong system. I recognized a related problem from my UIC fall internship project: an HR conversational agentic assistant with prompts, intent definitions and exceptions scattered across files.

There were two problems. The assistant loaded thousands of tokens of unrelated instructions, while engineers had to untangle hundreds of lines of repeated rules and buried exceptions before changing anything safely. We worked through that cleanup manually. Promptimize makes the process reviewable and repeatable SEAMLESSLY.

Upload a supported promptbase and optional project documentation. Meta Muse Spark extracts source-linked rules and their relationships. Review the proposed workflow folders, edit names and corrections, then freeze the candidate. A loader selects relevant instructions, includes connected exceptions and expands context when needed.

Run the same reviewed scenarios against the original and candidate. Visual traces compare model decisions, simulated actions and final states, alongside tokens, estimated cost and timing. Failures become findings for the next proposal; they never silently rewrite the frozen candidate.

Download real workflow folders, prompts, JSON intent registries, source evidence and a Python loader. An optional Spark/Wonder architecture advisor proposes loading techniques and starter adapters.

The stack is HTML/CSS/JavaScript, Node.js API functions on Vercel, the Meta Model API and an exported Python runtime.

In Forma, our fictional HR assistant mockup app that I built myself based on my internship project, both versions passed 16/16 checks across eight scenarios repeated twice. Total task tokens fell 58.5% and estimated task cost fell 51.9%, excluding one-time optimization. Latency did not improve. These are bounded demonstration results, not a production-wide accuracy claim.

SEE THE GITHUB for proof of performance comparison: https://github.com/stormragemc/Promptimize/blob/main/docs/FORMA_SHOWCASE.md

raw evidence: https://github.com/stormragemc/Promptimize/blob/main/demo/generated/benchmark-cloud.json

## More from Meta Global AI Developer Hackathon

- [SmalsaLab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2621)
- [Danna Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2622)
- [PolicyProof](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2623)
- [Save floods](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2625)
- [Maya's era](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2626)
- [Along](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2627)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2624. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt