Overarching project PRD before feature specs

[RPainter8West]

I'm finding spec-kit useful for feature implementation. But I'm wondering how it fits into a greenfield project where I want to create the overall PRD. Can it be used as is, and maybe rename the spec/001-MyFeature/ folder to something like spec/000-PRD/

Or should this be a separate mode with different prompts similar to /spcify and /clarify ?

I would like to get my project managers storing this history in my source repository.

Or is there a better tool for this, that I can take the output from and plonk it into my repo as a starting point. If so, what does this need to look like for spec-kit to honor it when planning the individual features?

[Zyzzx]

Before SpecKit, using other tools like kikocode/roo architect or orchestration modes, I have written up specs or PRDs and instructed it to use that as a guide to plan out building something. It proved useful first to get my thoughts down and organize everything in my head, and also provide the basis for a repeatable process that I can refine.

I am thinking more about what I can do as a developer and an architect and what the LLMs can do to actually increase productivity or quality. What I am seeing is that people want to offload as much as they can to the AI tools, falling into a trap almost where the LLMs are perceived to be more capable than they actually are, holding on to a vague hope that they will get "better".

The purpose of these tools isn't to live out some comic book or sci-fi fantasy, it's to accomplish real work, reducing burdens. The more we do up front, and the more succinct we map out a project, perhaps the more successful we will be. Creating PRDs and specs up front before we even touch tools allows a much more robust process for evaluating different models and agents since each time we drop the same prompt to generate a constitution, spec or plan with a different model, its going to be different every time.

I am approaching it also from the point of view that context is a scarce resource and tokens are an expensive commodity. There are limitations and failings of LLMs that need to be considered and treated as an insidious constraint. Specs and PRDs should be as terse and simple as possible, defining structures, patterns and methods without a lot of fluff that one might put into a document meant for business people, and leaning into what the LLMs are trained on, "understand" even.

[RPainter8West]

I think where the AI is helpful here is in finding things you may have overlooked in your manual PRD. Making sure all aspects have been covered, and there are no conflicting requirements.