HN Showcase
Back to browse
GitHub Repository

retrospec reverse-engineers plausible high-level spec prompts from Git commits using iterative Copilot SDK agent loops and similarity/realism scoring.

2 starsGo

Retrospec: reverse-engineer a spec prompt for an AI agent from a commit

by igolaizola·Feb 13, 2026·1 point·0 comments
Visit ProjectView on HN

AI Analysis

●●SolidBig BrainNiche Gem
The Take

It doesn't guess diff hunks — it runs iterative Copilot-agent loops and ranks candidate prompts by technical similarity and a separate 'realism' score, with explicit rules (no code blocks, structured markdown sections) to keep outputs human-like. The alpha-weighted scoring, model override, and prebuilt binaries show this is more than an experiment: it's practical for mining realistic specs from history or auditing intent at scale.

Target Audience

Backend/frontend developers, engineering managers, researchers building prompt->code datasets, and anyone wanting to infer intent from commits

Post Description

Hi HN, I built (vibecoded) Retrospec, a commit-to-prompt tool.

Given a repo + a specific commit, it iteratively searches for a plausible high-level spec prompt that could have produced that change. It runs agent loops, scores candidates for technical similarity and "realism" (does this look like a prompt a human would actually write), and outputs the best spec.

Inspiration: I saw Mitchell Hashimoto mention experimenting with agents to reproduce manual code edits, and around the same time GitHub released the Copilot SDK.

Repo: github.com/igolaizola/retrospec

Similar Projects

Developer Tools●●Solid

SpecFact CLI – reverse engineer Python repos into enforceable specs

This repo actually wires the specfact CLI to a tiny, reproducible codebase so you can import-from-code, generate .specfact bundles, and run enforcement presets with one-liners. The backlog-sync adapter and a deliberately buggy sidecar demo make failure modes easy to exercise, and the README lists exact smoke commands and logs to verify results. Inferred specs will always risk false positives, but the project shows practical artifacts (change_tracking, results logs) rather than theory.

Niche GemSolve My Problem
noldai
113mo ago
AI/ML●●Solid

Agents shouldn't operate software–they should coordinate commitments

Shifts agents from tool operators to commitment coordinators with evidence settlement.

Bold BetBig Brain
csehammad
303mo ago