Implement only does everything at once?

[romanr]

I read every document but it's still unclear; the implement command does it take task as parameter or only can do all tasks?

If only can do all - what about projects that are mostly complete? It took description of such project and created tasks for building it from scratch. what am I supposed to do, mark those tasks completed?

How do you work incrementally anyway?

[lioninlife]

I have the same question.

[brianluby]

I ran into this as well. I did find I could go back and point out it didn't implement everything. E.g. /speckit.imlplement you built some of the items but task-017, task018 etc were not completed. The first time it happened I just had gemini do the tasks of the markdown file but that lost me the speckit constitution and such. I am about to tackle a significant spec that will end up with lots of tasks hence my showing up here after my last large task. I am going to try claude opus 4.5 because I have had excellent success in creating additional use cases for the SPEC. But I am worried there is some upper limit where we need to break things into smaller chunks. So Spec Driven development but with an Agile implementation. That't the thing we are starting to move faster and faster so we are definitely finding edge cases.

Hope this helps.

[anchildress1]

Docs for each of these prompts is next on my list! 😄

To answer your questions, the /speckit.implement prompt is designed to implement tasks sequentially. So if you're not explicit in your prompt, then the agent will choose the next outstanding task in your list and is likely to keep going until it hits a system instruction or context window that tells it to stop. That stopping point depends on which agent you're using, though.

If you're looking for the agent to implement only the first task and then stop for review, then you can specify that in your prompt using something like this, for example:

/speckit.implement phase 1

Each agent handles things a bit differently, so if this doesn't get you the right result try rephrasing slightly. It should work everywhere, but LLMs can be picky sometimes! Also for the record, you're likely to have issues if you prompt out of order—I'd have to test it more myself to know for sure.