Metaprompting for Large Projects: A Deep Dive into the GSD Framework
How to streamline your spec-driven development using parallel agents, AI peer reviews, and autonomous execution.
After months of agent development, I’ve grown to love two metaprompting frameworks for Claude Code. In this post, I will describe the first one in more detail. I’ll also include three commands that have been a massive personal time-saver for me.
GSD - Get Sh** Done
This is the most structured framework for large projects. It’s ideal for spec-driven development, and working with it closely resembles the day-to-day workflow of an agile development team.
From the basic project definition, you move forward through milestones (epics) and phases (stories).
The entire framework is research- and token-heavy. It utilizes 4 parallel agents to research a given topic. Your specification is potentially supplemented with further research findings, resulting in a finalized agreement on requirements and a roadmap.
Each phase then has its own planning and execution.
In the meantime, there can also be an agreement on the UI, additional research during the planning phase, or a discussion of the plan itself. Again, I see a huge parallel with refinement sessions in an agile team.
Compared to other frameworks, GSD also features excellent auditing and milestone completion.
This ensures that no requirement slips through the cracks or goes unfulfilled. It also keeps an exemplary document tracking any potential technical debt.
Here are three tricks that I find absolutely awesome and that make the work so much easier:
1. /gsd:list-phase-assumptions
Personally, I consider this an absolute game-changer for my work.
I’ve worked in IT for ten years, seven of which I spent assigning work to developers from various teams across different banks.
I often assigned work to people from other teams whom I didn’t know very well. It was quite complicated to hand over complex requirements and simultaneously get feedback to ensure they understood it correctly.
Organizations all over the world struggle with this. It is the most common bottleneck where software development gets delayed due to mutual dependencies between teams.
With this command, Claude will list all of the explicit and implicit assumptions and risks that come to mind for a given phase.
If you were dealing with a real person, you’d struggle to pull three or four of these out of them during a half-hour meeting. Claude will list 20 assumptions driving its reasoning in under two minutes. It is extremely insightful, and it gives you the perfect opportunity to make corrections.
2. /gsd:review --phase XY --all (Peer review plans with external AIs)
If you have the Gemini CLI and Codex CLI installed alongside Claude Code, you can use this command to prompt Claude to consult its plan with other AIs.
The top-tier Codex and Gemini models will thoroughly review the given plan and compare it against the assumptions and requirements.
I only use this for the most complicated parts of development, and I’ve already had instances where these models caught something Claude hadn’t yet considered. This ultimately improved the quality of the plan. You sacrifice 3 minutes for a consultation but save hours of rework.
3. /gsd:autonomous
Autonomous mode can power through all the remaining planning and execution phases according to the roadmap entirely on its own.
I use this when a given milestone is relatively simple and the requirements are unambiguous. It’s really pointless to sit at the PC just clicking Yes, Yes, Yes. Finally, there is a mode that handles it by itself.
If you are on the go, you can monitor Claude’s progress via remote control in the Claude App.
I actually keep remote control running all the time now. You just need to enable a specific flag in the config, and your Claude Code threads will automatically appear in the Code tab of the mobile app.
I hope this proves useful to you. I’m at a stage where I’ve progressed heavily in my AI usage this year compared to last year’s experiments. At the same time, I am constantly discovering and exploring new workflows myself.
I’d love it if you shared your best practices and tricks for getting high-quality results with me in the comments or via messages!
PS: Link to the GSD is here. Share this articles with your friends struggling with AI development.
Interesting! I'm a web developer and only just now really kicking off with Claude's ecosystem. This is invaluable.