How I Keep AI-Built Projects Under Control
I use OpenGSD for almost all my projects. It costs more tokens, but it makes the specification explicit before the code is written.
Building with an AI coding agent goes better when the plan is written down first. That is why I use a planning workflow.
What I use
I use OpenGSD for almost all my work. It structures a project into written pieces before and during the build.
How it works in practice
- A project file sets out what I am building and why
- Requirements describe what done looks like
- A roadmap breaks the work into phases
- Each phase gets a plan, which is then carried out and verified
Claude Code does the building, and I check what it produces as it goes.
The cost
It uses a lot more tokens than asking an agent to just build something. With many projects running on Claude's 20x plan, I still hit limits. I accept that because the clearer specification saves me time sorting out confusion later.
What I still do myself
The workflow does not replace checking. I read the changes, test the result, and ask the agent to check and secure things properly. I also think AI-written code tends to bloat, so I expect to simplify some of it later.
What I would tell someone starting
Write down what you want before the agent starts. Check everything it produces. And keep your own judgement involved at every step.
Comments