FIRST IMPRESSIONS ON GITHUB SPEC KIT

I came across GitHub Spec-kit and I decided to give it a try. It looked like a lighter version of the BMAD Method. Not that I dislike BMAD but in my quest to finding the right methodology for myself, I kind-of want something lighter than BMAD and Spec-Kit kind-of look like it.

My first impressions are not negative nor positive yet. It is similar to BMAD on both the positives and the negatives so far.

Positives

• It creates only 3 slash commands: /specify, /plan, /task.
  • First for defining the features from the business standpoint, second to define the high-level technical implementation plan, and third one to break down the plan into tasks.
  • No research phase, no interaction with agents to define each part of the process, it’s pretty much you telling it what to do and how to do it (high-level).
• It uses shell scripts for common repetitive (create a new branch for the new feature, etc.) tasks which I’m sure saves a few tokens and makes the process a bit more deterministic.
• It also uses predefined templates for the different spec files and includes them in your project root folder. This gives you freedom to modify whatever you want for your project.
• I love the creation of a new branch for a new feature.
• Process is Test-Driven development, which forces you to look into unit testing very early.

Negatives

• With BMAD it was more explicit the handoff between the different agents. One takes you to the next one. With Spec-Kit, you are slightly more on your own. You need to run the custom slash commands at the right time.
• Piggiback’ing on the previous one, there is a checklist on the docs generated by Spec-kit but it is unclear who the owner is. Is it me? Should it be the LLM after running the tasks? The official documentation is not that clear, it shows an example where the user gives the instruction to validates one of the checklists but then there are other checklists.
  • I needed to explicitly tell the model to run the checklists. It would have been nice that the process automatically check those that are related to its own process (e.g., Phase 3: Tasks generated (/tasks command)).
• Purpose of some filenames is confusing because of the name. It is only after you read the contents when you understand. Examples: research.md, quickstart.md.
• For some reason, it feels slower than BMAD. Maybe because with BMAD there is a big effort upfront, then the iteration of implementation feels faster. With Spec-kit that big upfront effort is replaced by small-to-medium efforts for each and every feature you want to implement. This is hard to compare and judge but my impression is that the actual implementation is slower than BMAD.
  • I definitely need to make more progress, this is just my first impression, but maybe later on, the process will speed up.

Conclusion

Too early to tell. I need to develop a few features using the process and then I will be able to provide a better feedback.