IdeaWell

Project ideas from Hacker News discussions.

Analyzing Frontier Model Progress with My Favourite Game: Prince of Persia

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Three prevalent themes in the discussion

  1. Prompt quality is a major obstacle
  2. “The prompts provided are atrocious. It's amazing that the LLMs actually built something useful.” – smokel
  3. “It's so confusing how the actual article is in English, but the prompts are just gibberish.” – InsideOutSanta
  4. “Yeah, I'm thinking the nigh‑unreadable AI-speak we get these days makes a lot more sense if this is what they're training on. Or maybe the author has translated the prompts from another language?” – daemonologist

  5. “Or having another model proofread the prompts and write clearer instructions.” – cavemandaveman

  6. Models benefit from starting fresh each iteration rather than building on flawed outputs

  7. “I feel like the LLMs would have done better if they had started from scratch each time, rather than being burdened by the output from the previous attempt.” – InsideOutSanta
  8. “Testing the next model by iterating on the first model’s crappy foundation. Why not start from the original each time?” – aksss

  9. “Instead of fixing a broken version, maybe each model should have started from scratch.” – tehlike

  10. Observations of LLM performance versus human‑crafted solutions (and hardware feasibility)

  11. “Astra … had a highscore of around 1.7M … it disassembled parts of the ROM to extract information about the game.” – criemen
  12. “Astra … reacts to the visuals on‑screen, planning ahead by estimating velocity of all objects on screen… behaves more like a regular 'perfect' player.” – criemen (comparing Astra to the winning human solution)
  13. “You could actually play Prince of Persia on an 8088 with a CGA card!” – hapless (highlighting that the original game runs on very low‑end hardware)
  14. “Prince of Persia is a piece of contemplative, subtle, beautifully, artistically minimal motion puzzle art.” – dofm (appreciation of the game’s design amid the technical discussion)

🚀 Project Ideas

PromptPolish: AI‑Powered Prompt Clarifier

Summary

  • Takes garbled or non‑English prompts and rewrites them into clear, concise instructions for LLMs.
  • Core value: improves LLM output quality by reducing prompt ambiguity, saving tokens and iteration time.

Details

Key Value
Target Audience Developers, prompt engineers, non‑native English speakers using LLMs for code generation or content creation
Core Feature AI‑driven prompt clarification and translation, with optional clarity scoring
Tech Stack React frontend, Python/FastAPI backend, leveraging GPT‑4o for rewriting and DeepL for translation
Difficulty Medium
Monetization Revenue-ready: subscription tier (e.g., $9/mo for 10k clarifications)

Notes

  • HN commenters lamented “The prompts provided are atrocious” and suggested “having another model proofread the prompts and write clearer instructions.”
  • Could spark discussion on prompt engineering best practices and reduce wasted compute on misguided LLM attempts.

BaselineBot: Prompt Versioning & Reset Service

Summary

  • Stores the original specification (code snippet, design doc) and forces each LLM iteration to start from that baseline, preventing error propagation.
  • Core value: ensures fair model comparison and reduces frustration from building on flawed AI‑generated foundations.

Details

Key Value
Target Audience AI‑assisted developers, researchers comparing LLMs, anyone iterating on LLM‑generated code
Core Feature Version‑controlled baseline storage with a “reset to baseline” button; optional diff of LLM output vs. baseline
Tech Stack Electron/VS Code extension, Git‑LFS or SQLite for storage, CLI wrappers around LLM APIs
Difficulty High
Monetization Revenue-ready: SaaS plan $15/mo per user, free tier for hobbyists

Notes

  • Users complained “testing the next model by iterating on the first model’s crappy foundation. Why not start from the original each time?” and noted LLMs would do better if they started from scratch.
  • Provides a concrete tool to implement that suggestion, likely to be embraced by HN crowd interested in reproducible AI experiments.

LinguaPrompt: Multilingual Prompt Translator & Optimizer

Summary

  • Detects the language of a user’s prompt, translates it to English (or target language) while preserving technical terminology, and then optimizes phrasing for LLM comprehension.
  • Core value: eliminates language barriers that cause LLMs to misinterpret prompts, especially for non‑English speakers.

Details

Key Value
Target Audience Global developers, educators, non‑English speaking teams using LLMs for coding, documentation, or brainstorming
Core Feature Automatic language detection, translation, and prompt rewriting with confidence score and syntax‑preserve toggle
Tech Stack Svelte frontend, Python backend, using LibreTranslate/Azure Translator for detection/translation and GPT‑4o for polishing
Difficulty Low-Medium
Monetization Revenue-ready: usage‑based pricing $0.001 per prompt

Notes

  • Commenters pointed out language barriers (“models would do better being prompted in the author’s main language”) and the confusion of gibberish prompts.
  • Offers a practical utility that could be discussed in threads about prompt engineering across languages and improve accessibility of LLMs worldwide.

Read Later