Spec Driven with a Backlog

AI is changing many things today, especially the way in which we build software. I wanted to share my current workflow that is built on things that I've found useful.

There was the term "Vibe Coding" coined by Andrej Karpathy, subsequently also a very good book by Steve Yegge and Gene Kim https://itrevolution.com/product/vibe-coding-book/. Which has been quite interesting but in my experience limited to proof of concept / idea validation.

But then I found the excellent project from Den Delimarsky for spec driven development https://github.com/github/spec-kit and then this added a bit of structure into the coding workflow. Which pretty much works with any Agentic Coding tool!

While attending the amazing AINativeDev summit hosted by Patrick Debois. I learned of Alex Gavrilescu's project https://github.com/MrLesk/Backlog.md which is a very good project to keep track of work right in the repo. (don't worry I don't think this changes need for planning tools like Atlassian Jira / GitHub Issues, etc) but it allows the task status + work produced to track together (and gives the Agent easier access to what it should be working on)

And then I started to experiment and put the two together. A spec driven approach (customizable to my needs) with a backlog of where we (the coding agents and I) are in our work, right in the git repository.

I am still evolving this, and tweaking the prompts and improving it, adding a bit of rigor to things, adding task specific memory, to help solve the "Context Engineering" problem that sometimes arises.

Here is a visual of my basic workflow. It can be edited in a web or terminal and pretty much use most coding agents. (Claude Code is my primary). I also often assign a different agent / model to validate the code, as I've found that seems to be a good balance.

If you hear "You are Absolutely Right!" you know things have gone wrong. I am sure some people enjoy the sycophantic agents, but I prefer brutally honest and objective feedback.

I hope you will consider trying this out, use it or fork it.. and give feedback (in github) raise an issue, suggest a feature. Its not going to get better without more people using it and giving feedback and sharing their workflows.