How are people automating execution?

[jcgottlieb]

Hey all! Assuming all spec-kit artifacts are in place and it's time to implement...

How are people automating execution from A to Z?

I'm hard stuck manually shepherding small enough batches of tasks that they don't blow out the context window and burn premium model cash, because I'm also stuck getting less than premium models to deliver first pass quality.

I'm chasing a repeatable end-to-end automation loop, even if it means adding orchestration tool$ to the stack.

Thanks in advance!

[anchildress1]

Personally, the setup is too expensive for me to warrant using it consistently enough to think about an automated flow. I love the project and it definitely makes a difference! It also easily doubles my AI cost running it effectively enough to make that sort of difference. So, I end up reserving it for only very complex tasks and the occasional hack project when I can make it work.

This repo makes very good use of the architecture that Copilot has built in, though. I haven't gotten close to documenting it yet, but if you init a project with --ai copilot then sub-agents are already created for you and are wired to work inside of VS Code. It's still a very manual process, but that's the only jump start that exists currently.

[unphased]

i like the idea of having enough structure around spec and documentation for a project to the extent of providing templates that we can use LLMs to also help us fill out, but to be honest this is what we've been doing for months (years) already and it works great without the guidance of such a framework. Whether we use spec-kit or raw dog generating the project guideline documentation, once you have that guidance in place in your project and your workflow automatically pulls them into context when you do assisted coding in those projects with LLM coding (agentic or otherwise) tools, you are more or less fully cooking at that point.

Since my workflow is not currently centralized around vs code, I am here looking for tooling that lets me get these "slash commands" in a spec-kit CLI, rather than having to fire up vs code just to use spec-kit efficiently.

Unfortunately, as far as I can tell there is no such thing. Someone please clarify.

[anchildress1]

Depends on whether or not your workflow utilizes Copilot or not. I just caught up with GitHub's integration of this custom agent idea earlier today, and apparently I stand corrected—it's not just VS Code anymore, although that's still where you'll find the best integration. IntelliJ, Eclipse, Xcode, Copilot CLI, and coding agent on GitHub.com all apparently support custom agents according to their docs.

That being said, if it were me with not-copilot branded agent setup abilities? I'd most definitely be using Spec Kit configured for a copilot agent just to create those baseline agent files. Then where they may or may not migrate to after that is anybody's guess 🤷🏻‍♀️ 😆 It would probably take a bit of customization and finagling to get it operable again, but it's not starting from scratch either!

[camiloborges]

Hey there.
I've been working on a somewhat largeish project with spec kit and I've added few things to an 'orchestration/main spec' repository.

• I create the spec for the changes across a number of repositories, drawn from a corporate solution design document (that was also produced with AI).
• a jira tracking document, that keeps track of all user stories that have been mapped to a 'central' spec.
• then, I have an 'speckit.orchestrate', that engineer uses to initiate the process for one of the stories in the initiative, by asking where the repository is, and copies speckit files + main constitution to repository, and produces a repository specific constitution (see that as initialising repo for speckit). this is done from one instance/window of cursor
• then, engineer opens target repository on another cursor window, then instructs agent to start guide for an user story, which describes what the story is about, describes the planned delivery, and ask about 'granularity' preferences, which adjusts plan.

This seems to work well for backend services, but requires specification done to a level that an offshore team would be able to deliver. Takes a bit of time to get to the standard I wanted, but I am really surprised by the quality of the output.
I am still to figure out a good flow for developing front end, I am having mixed, but improving results when working with a pre established front end.
I am using cursor over vscode because I can easily ask cursor to make changes to a folder that isn't within my IDE window context. You need to add stuff to workspaces in vscode, but that doesn't work when people have repositories on different folders.
I am planning/doing deep research with opus 4.5, and executing/implementing with composer or sonner 4.5