· 1 min read
GSD: how I got AI agents to actually finish the work
The GSD (Git. Ship. Done) framework, and the /goal trick in Claude Code that keeps the agent working through my whole plan without stopping every few minutes.
Translated from the original Arabic post.
Since I started vibe coding, my biggest problem has been AI laziness, and Claude Code in particular 🤦♀️
I’d write a complete requirements file and hand it over, and it would build something that reflected maybe 10% of what I asked for, then tell me it was done. On top of that, every time I asked for something, it would stall and tell me to do it myself.
Then I stumbled on a framework that saved me from all that effort of convincing a lazy AI to deliver what I want.
GSD (Git. Ship. Done) 🚀
It’s a framework for AI agents that organizes the work into phases. How many phases you have depends on the size of the work and how much you care about its details. Each phase runs the same loop:
- Discuss: it talks through the details of the work with you, so it can draw up the right plan.
- Plan: it researches and double-checks the plan.
- Execute: it starts carrying out the plans, in parallel.
- Verify: it checks what was built and fixes the problems.
- Ship: it opens a pull request, archives the phase, and moves on to the next one.
Another nice touch: you can pick the right model for each step. For example, if you want Opus 4.8 for planning but Sonnet 4.6 for execution, that’s easy to set up.
The stopping problem, and /goal
The next problem I hit was it stopping so I could clear the context or approve things. That’s where some Claude Code features come in, like /goal: I set the goal to finishing all phases, and it keeps going through my plan unless it needs input or a confirmation from me.
I hope my experience helps. If you have other tips, please share them.