IdeaWell

Project ideas from Hacker News discussions.

Show HN: Parsimony – an open-source benchmark for how much code AI models add

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Theme 1: Measuring and reducing LLM‑generated verbosity
- “On the one hand, I absolutely agree that we should try to measure and reduce verbosity that LLMs are slopping into our codebases.” – fastball

Theme 2: Trade‑off between brevity and readability
- “Astra plays code golf… My own revealed preference is that I'd rather have the excessive comments and tests of Claude rather than the inscrutability of Astra.” – fastball

Theme 3: Preference for explicit, well‑documented code over terse “golfed” output
- “I'd rather have the excessive comments and tests of Claude rather than the inscrutability of Astra.” – fastball

🚀 Project Ideas

LLM Code Styler

Summary

  • Automatically post‑processes LLM‑generated source code to add readable comments, docstrings, and basic unit‑test scaffolding while preserving functionality.
  • Core value proposition: gives developers the concise logic of models like Astra with the maintainability and clarity of Claude‑style output.

Details

Key Value
Target Audience Software engineers and teams using LLMs for code generation who want readable, maintainable output
Core Feature Rule‑based and LLM‑assisted transformer that inserts comments, extracts logic into functions, and generates Jest/PyTest stubs
Tech Stack Python (ast, libclang), optional LLM fine‑tuning via HuggingFace, CLI + VS Code extension
Difficulty Medium
Monetization Hobby

Notes

  • HN commenters lamented Astra’s “unreadable code” and preferred Claude’s excessive comments/tests; this tool directly addresses that trade‑off.
  • Enables teams to keep using powerful LLMs while enforcing internal style guides, reducing review friction.

Verbosity Gauge

Summary

  • Provides a real‑time dashboard that quantifies code verbosity (lines of code, comment density, test‑to‑code ratio, cyclomatic complexity) for LLM‑produced snippets.
  • Core value proposition: lets teams set and enforce verbosity thresholds, making the trade‑off between conciseness and readability explicit and measurable.

Details

Key Value
Target Audience Tech leads, code reviewers, and DevOps engineers overseeing LLM‑assisted development pipelines
Core Feature Metrics collection plugin (GitHub Action, pre‑commit hook) + web UI showing trends and alerts when verbosity falls outside configured bands
Tech Stack Go (backend), React (frontend), Prometheus/Grafana for metrics, Docker for deployment
Difficulty Medium
Monetization Revenue-ready: SaaS subscription per active repository

Notes

  • Users complained that Astra “wins” benchmarks by code golf but sacrifices readability; Verbosity Gauge makes that loss visible.
  • Facilitates data‑driven discussions about acceptable verbosity, aligning with the HN thread’s desire for measurable standards.

Prompt Verbosity Tuner

Summary

  • A prompt‑engineering helper that suggests modifications to LLM prompts to steer output toward a desired verbosity level (more comments/tests or more concise).
  • Core value proposition: reduces trial‑and‑error prompting, letting developers reliably get the style they prefer from any LLM backend.

Details

Key Value
Target Audience Developers who craft prompts for LLMs (e.g., via Copilot, ChatGPT API, open‑source models)
Core Feature Interactive UI that analyses prompt text, recommends additions like “please include docstrings” or “keep under 30 lines”, and predicts resulting verbosity via a lightweight classifier
Tech Stack TypeScript (React), TensorFlow.js or ONNX for classifier, optional backend via FastAPI for model serving
Difficulty Low
Monetization Hobby

Notes

  • Commenters expressed a preference for Claude’s verbosity but found Astra’s terse output “unreadable”; this tool lets them guide LLMs toward their preferred style without switching models.
  • Encourages experimentation and sharing of effective prompt patterns, which could spark further discussion on HN.

Read Later