Let me introduce my "Micro spec agentic flow".

👉 https://gist.github.com/blambeau/ba23eb9d74cf42b3ca8a210d1567cad6

I very rarely ask AI agents to work on big upfront plans.
It simply does not work, for the same reason it does not work with humans: too much uncertainty.
Instead, I pair program with Claude Code using baby steps, in a setup close to eXtreme Programming (XP).

Each step is a small card, whose development takes one hour MAX and can be committed to main.
Tasks and their changes are part of the git history (backlog in code).

The agent instructions start like this (see whole version on GitHub):

# Method: micro spec agentic flow

We follow a micro spec approach where each development task is described as a short markdown card annotated with data in YAML front matter.

## Workflow rules

* ALWAYS work one card at a time.
* ALWAYS finish the ongoing card(s) before working on another one.
* If you plan something, write your plan to the card and mark it `analyzed`.
* If you move a card from `todo` to `ongoing`, make sure to include your plan and to-do list within the card.
* If you finish implementing something, write a summary of changes to the card and mark it `done`.
* If you've been prompted to work on something without having a `todo` or `ongoing` card, create a new card **before** starting the work.
* A card MUST exist and be included in every commit. Never commit without the corresponding card file.
* Every time you reach a milestone, commit everything, including changes to the task file(s).
* Every time you commit, run the tests first, they **must** pass.

[... see on GitHub, link above ...]

It's simple, it works.
It can be put in CLAUDE.md or as a SKILL.md.

There's nothing really new here.
That's simply how I work for years.
I'm impressed by how effective this method remains in the age of AI.

#SoftwareEngineering #Agile #IA #ClaudeCode