HN Showcase
Back to browse
Gravel – allow non-engineers to make prompt PRs

Gravel – allow non-engineers to make prompt PRs

by paraknight·May 26, 2026·1 point·0 comments
Visit ProjectView on HN

AI Analysis

●●SolidSolve My ProblemSlick

Git-backed prompt editor lets clinicians make PRs without knowing git.

Strengths
  • •Prompts live in git as single source of truth alongside application code
  • •Production trace review creates prompt change suggestions and evals automatically
  • •llms.txt install contract for AI agents shows thoughtful developer experience
Weaknesses
  • •Emerging category means unproven workflow adoption patterns
  • •Requires embedding into existing apps which adds integration overhead
Target Audience

Vertical AI startups, teams building agents with domain experts

Similar To

Langfuse · PromptLayer · Arize

Post Description

Hi HN, I'm Yousef. I've built a lot of agents for vertical AI startups in the past few years, and I've become convinced that prompts are code, so must be kept in git.

The startups that decouple do so because engineers build agents with git, and non-engineers (lawyers/clinicians/etc) align agents outside of git.

So the problem of coordination and versioning shifts to out-of-sync docs and platforms.

This is why I built Gravel.

An open-source library that embeds a prompt editor straight into your app. Your prompts have a single source of truth (your codebase).

After a minute of setup:

1. Your team can make prompt PRs without knowing what a PR is

2. They can review outputs straight from your production traces

3. Those reviews create prompt change suggestions and evals

If you're building vertical agents, would love your feedback on it!

Similar Projects

AI/ML●●Solid

Prompt Builder – A block-based editor for composing AI prompts

The block metaphor and live compiled preview are honest, practical improvements for anyone wrestling with long, conditional prompts — toggles for A/B testing and global {{vars}} are especially handy. Multi-model execution and editable response panes show the author thought about iteration and comparison, but the screenshot feels safe and functional rather than boldly new; I want to know how it handles collaboration, exports, and model/credit management.

Solve My ProblemNiche Gem
Jaber_Said
103mo ago