HN Showcase
Back to browse
GitHub Repository
6 starsTypeScript

spec2commit – I automated my Claude Code and Codex workflow

by arbayi·Feb 21, 2026·2 points·0 comments
Visit ProjectView on HN

AI Analysis

●MidShip It

Orchestrates two AI tools you already pay for; requires both Codex and Claude.

Strengths
  • •Honest about the workflow it automates—removes manual copy-paste between two specific tools.
  • •State persistence across sessions lets you pause and resume mid-pipeline.
  • •Review loop retry logic (up to 10 iterations) adds real feedback cycles instead of one-shot generation.
Weaknesses
  • •Depends on two separate, expensive proprietary tools (Claude Code + Codex); locks in to OpenAI ecosystem.
  • •Zero evidence of actual output quality or adoption; no demo repos, benchmarks, or success stories shared.
Target Audience

Solo/small team developers juggling multiple projects who already use Claude Code and OpenAI Codex

Similar To

Aider · Continue IDE plugin · GitHub Copilot Workspace

Post Description

I usually juggle multiple projects at once. One is always the priority, production work, but I like keeping side projects moving too.

My typical flow for those back burner projects was something like this. I would chat with Codex to figure out what to build next, we would shape it into a Jira style task, then I would hand that to Claude Code to make a plan. Then I would ask Codex to review the plan, go back and forth until we were both happy, then Claude Code would implement it, Codex would review the code, and we would repeat until it felt solid.

I genuinely find Codex useful for debugging and code review. Claude Code feels better to me for the actual coding. So I wanted to keep both in the loop, just without me being the one passing things between them manually.

My first instinct was to get the two tools talking to each other directly. I looked into using Codex as an MCP server inside Claude Code but it didn't work the way I hoped. I also tried Claude Code hooks but that didn't pan out either. So I ended up going with chained CLI calls. Both tools support sessions so this turned out to be the cleanest option.

The result is spec2commit. You chat with Codex to define what you want to build, type /go, and the rest runs on its own. Claude plans and codes, Codex reviews, they loop until things are solid or you step in.

This was what I needed on side projects that don't need my full attention. Sharing in case anyone else is working with a similar setup.

GitHub: https://github.com/baturyilmaz/spec2commit

Similar Projects

Developer Tools●●Solid

Turn Claude Code or Codex into proactive, autonomous 24/7 AI agents

Turns an existing Claude/OpenAI key into something that actually keeps working for you: heartbeat pulses propose tasks, persistent research notes and daily journals build context over time, and browser control plus messaging (Telegram/WhatsApp/Slack) make it proactively useful. Clever local-first architecture and a polished desktop UI give it real stickiness, but it's boxed in by macOS-only delivery and the need for paid third-party models.

Niche GemWizardrySlick
alternateman
323mo ago
Developer Tools●●●Banger

Oneprompt – Problem description to MVP (experimental)

This isn't just prompt chaining — it's a complete methodology baked into a Claude skill that actually closes the loop. The clever part: personas created during early research come back to usability-test the built app before launch, surfacing real UX friction that pure iteration never catches. It generates working Next.js apps with documented artifacts at every stage, which means you get both a product *and* the research trail proving it solves a real problem.

Big BrainZero to OneBold Bet
LambdaAndLatte
203mo ago