Is it ok to provide a complete feature spec for the /specify prompt?

[lance2k]

I use this Idea Generator agent https://github.com/github/awesome-copilot/blob/main/agents/simple-app-idea-generator.agent.md, to help me brainstorm and generate ideas. The agent will return a spec at the end and I used this as input to /specify prompt.

Am I doing the right thing or am I over-engineering things? From what I understand /specify prompt will create the spec.md which will contain the spec.

It would be nice if we have a /brainstorm prompt which will improve our idea/prompt before handing off to /specify prompt.

[itinsecurity]

Use whatever tool you want to generate a prompt for /specify. I spar with claude sometimes for help figuring out how to describe what I need/want - exploring options and variants of the ideas I provide. When I eventually pick the ideas I want, I specifically ask it to create a short/concise natural language description (instead of a structured prompt - that's what spec-kit provides). I have also created some custom commands to help me with things I do around spec-kit - you can just create your own custom /brainstorm command if you wish.
BUT... remember a few key points about the models, like: they are chatty, and will respond with a lot of nice looking words. Could cost you tokens and bloat your context if you don't strip your spec of fluff. They will also always try to agree with you and prioritise what they think you like to hear (which isn't always the same as "what is a good way to do what you want"). Again: you need to reality check and strip it down. AI generated ideas can quickly become overengineered. And bundling too much work into one spec at a time (if you're not on some large/enterprise plan) can quickly push your agent beyond your context window, creating all sorts of trouble.

Lastly, structuring your brainstorming too much could go against the whole point .... subtly forcing idea generation to "regress to the mean" so to speak. The LLM isn't imaginative, it's actually pretty boringly average.

[lance2k]

/specify provided me with 220 lines of spec, is this too long for a spec? I thought speckit will give me a close to MVP result, but it was far from it with alot of things missing from the user stories and errors in the code. It was a fullstack App though, Next.JS + Supabase + Shadcn UI. If I could go back, I would have set-up Next.js + Supabase manually because copilot messed up Tailwind4 and uses Next 15 instead of Next 16 even though I have custom instructions for Next 16.

[mnriem]

How well Spec Kit will perform depends on various factors. One indeed is the size of the specification. Note however that the underlying model used by your agent is also a large factor on what will and can be delivered.

[itinsecurity]

If you use the same model for that initial spec, how much better will it be the second time around? It might just end up chasing it's own tail. I'd try experimenting with giving /specify a shorter, natural language description, and see if the model makes do with following whatever restrictions are given elsewhere (like the constitution).
And if a "lot of things" are missing from an MVP, are you sure that's really a minimum product? The spec might not be too long, but perhaps it contains too much? A detailed description of one thing is different from a long list of things.

[ahernsean]

I recently gave /specify a 285-line MD document that described in significant detail what I want. Spec-kit produced an even more detailed spec.md with other supporting docs. I don't see anything wrong with the workflow you're proposing.

The AIs are good enough that you could probably brainstorm with one without needing a /brainstorm prompt. Since brainstorming has no particular needs from the later spec-kit artifacts, I don't see what value /brainstorm would bring.