Just implemented a feature using GitHub spec-kit
Just implemented a feature using GitHub spec-kit.
It's an opinionated workflow for coding via an LLM. I used it with Claude Code, but it works with many different LLMs.
Wow. It does a LOT.
- Creates feature branch
- Generates a high-level plan
- Creates a quickstart doc for trying the feature
- Generates research docs about decisions and alternatives
- Outputs a detailed spec
- Creates a list of tasks based on the spec
- Implements tests first and enforces that they fail (TDD)
- Does the implementation using task list.
- Validates it all works by running linting, build, and tests (it fixed its own minor mistakes at the end too).
It's VERY thorough and organized, and walks you through a highly opinionated development flow.
I told it to create an admin feature for my restaurant app and it generated and implemented these 32 tasks, checking them off as it progressed.
If you're coding with AI assistance, you should absolutely try this. It's a real eye-opener to how far you can take things.
Also, a few downsides:
1. Generates a LOT of files. Feels excessive to me. Certainly the opposite of lazily Vibe coding.
2. Uses a LOT of tokens, because it's very verbose, detailed, and specific.
3. If I change my mind midway through the flow, there's a LOT of files to update, and I'm unclear how I'd do that easily. Perhaps the LLM could figure it out...
4. Since there are so many files generated, it's easy to get lazy and just approve. This replaces the work of creating specs and writing code, with the work of reviewing specs, and reviewing code.
5. Despite all this, my feature didn't work quite right. So now I'm unsure if I should update the specs, or just use fix the issues myself with Claude's help.
Here's the repo: https://github.com/github/spec-kit