Am i missing something

[colinrees]

So.. I love the concept of spec kit, its clean non opinionated and you can see how it works...
But...

I am really struggling to get it to produce anything more than a simple static app, I have tried doing it all in one go, I have tried bootstrapping and then adding feature by feature.

Can I ask, have others had any success in getting it to help with more dynamic sites (I am using Django) and if so how have they approached it.

Best

[czei]

If you haven't tried spec-driven development before, then first you have to practice using that method of development where you are a technical manager, not a coder. This kit simply gives some nice templates and a framework that would take months to development on your own. This way of working means spending around half your time, if not more, on up-front planning. You not only have to make plans and tasks, but have a second AI review the plans and tasks that were fleshed out by the first LLM. Its no different that writing a spec for a big project and having it reviewed by another team before starting. And then after each sprint I have another AI do a code review of everything before moving on, again just like managing a team of junior programmers. So far I've done two large coding rewrites that would have taken a team of programmers 4-6 months, and done each in about a month. One was a major refactoring of an old Java app with more than 100,000 lines of original code and 50+ external libraries, taking it from Java 8 to Java 17, and from a 5,000 line ant build script to maven.

[AdventureBear]

what do you mean when you say have another AI review the plan? Is this with the commands created/ included in spec-kit? Or do you have other scripts you are referring to when you say "another ai"?

[czei]

spec-kit creates its files in specs/project-name. Once the plan.md and task.md files are created, I have another AI do a review of how well the plan matches the spec.md. And then each time the /implementation stops at a logical point I have another AI do a code review of the work. That's just the way I work, as it tends to keep hallucinations in check, and leads to better overall code quality. I've got subscriptions to claude, openai, gemini, and grok. Claude code is used as the orchestrator, and the others will help with debugging, spec reviews, code reviews, etc. Note that with larger projects you end up working in sprints, just like with a real team, so none of this is a large project done in one shot. You still have to manage the AI team, only the project goes much faster.

[jeremyeder]

There are also budgets to consider. We will route valuable queries accordingly. putting higher spend just prior to any human interaction. Definitely agree on the vendor-spread. Same as hyperscalers...