Once again, I spent too much time building an AI software development workflow.
Why#
I tried Claude Code. Codex. Pi. Kiro. Kimi Code. Copilot. Antigravity. Cursor. OpenCode. Superpowers. Ponytail.
And while we can feel the models getting better and better, I couldn’t help but hate the “standard” LLM-driven workflow:
- Prompt the coding agent about a feature
- Wait for it to gather context and respond
- Start iterating in the chat, respond to questions one by one, with a shitty interface
- Once the feature is well specced, let the agent build
- Be disappointed by the resulting code: too complex, faulty, poorly organized
- Add more skills, prompts, context, and rules, and pray it goes well
So I sat down, and just like I did for GHOST, I thought about how I wanted to
interact with an LLM:
- It needed to be asynchronous. I don’t want to sit down for 20 minutes while an LLM runs tests.
- It needed to be efficient. What I hate over everything is having to go back and play Q&A after an LLM has spewed a 10k LoC PR. I want good code from the start, following my specs, without any uncertainety.
- It needed to capture intent. If I was to not look at the code anymore, at least my decisions and directions should be stored in the repository, since those are what actually generated the code.
- It needed to be readable. Walls of text in TUIs don’t do it for me. I want context, hyperlinks, graphs, formatting, and visual hierarchy.
- It needed to be robust. LLMs change every week. Harnesses too. It should be a workflow that’s understandable from just reading text files in the repo.
- It needed to work on my phone. I don’t want to be glued to my PC, waiting for an agent to reply to me. I want to be able to evaluate specs and features on the go.
All those things also would have the advantage of working particularly well with local models:
- Even if running a model is a bit slower locally, it’s ok if it’s asynchronous
- Being strongly opinionated directs agents in the right direction
What#
So that’s how I ended up creating uraeda:
It’s a git-driven workflow. All stable state gets comitted to git:
- First, I create a proposal explaining the big picture of the feature or change I
want to operate
proposal.md
- The agent works autonomously, in-depth, and researches the current state of the
codebase, possible issues with the feature, and how it’s implemented in other projects
discovery.md
- If needed (it often is), the user asks for clarifications: a list of questions to
clarify anything that’s unclear about the proposal and take important engineering
decisions. All with necessary context.
clarifications.md
- After responding to the clarifications, the agent creates a spec. It includes user
stories, important decisions to save, edge cases, mockups, …
spec.md- decisions/[topic].md
- At this point, I will have a few comments, and I’ll ask for a v2 of the spec
spec-review.md
- Once the spec is validated, the agent creates a step-by-step implementation plan,
highlighting waves of tasks that can be developped in parallel.
plan.md
- After the implementation, the agent issues a PR and condenses the information from all the flow in a single report, committed to the repository.
The decisions and final report are the only artifacts that stay in the repo, as they will give context beyond code for the future agents.
Why the why?#
It’s complicated#
I still have a love-hate relationship with LLMs and their impact on coding.
On the one hand, I finally have been able to build software I’ve always wanted for myself. I have tons of vibe-coded apps I use daily, from my own restaurant recommandation system to bicycle parking spot search engine. Tons of minimal, extremely fast, simple to use micro-apps. And I can also envision building bigger, better apps than I ever did.
I can even produce frontends I’ve always imagined but was never able to write as I hate Javascript!
On the other hand, all the feeling of accomplishment, success, and flow state I had when writing code is gone.
I enjoy seeing the final result, and I enjoy having great, tailor-made software.
But I hate what it’s done to me. I used to write Rust serverless functions with sub 10ms startup time. I used to have 100% code coverage on the backends I developed at my company, all by hand. I used to produce high quality code that people loved working with.
It’s all gone. I don’t even know what the code looks like anymore. Half of it is written in languages I don’t even enjoy, and wouldn’t be able to evaluate since I didn’t have time to develop a strong taste for it. I’ve become a project manager with a strong technical background.