Anthropic recently released a guide on the AI-native SDLC, so I did what I do with most guides I come across: I put it next to my own workflow, the one I posted about recently.
When a guide comes from the company that builds the models, we have a tendency to take it as is and just apply it. What I do instead is go item by item, brainstorming with the agent on what works for me and what doesn't: "this one yes, this one no...", and so on. Then I test the ones that made the cut.
Some of it was already there: plan -> build -> review was very aligned, while some other practices were not.
The key gap was testing. Claude was already testing its work proactively, it just wasn't written anywhere in my workflow. The other change was a Proof section in every plan, a short note on how we'll know the thing works. As a PM I read the plan, not the diff, so that's where I want to see it.
What I left out, for now: extra files and automatic blocks on what the model can do. I don't know yet if I need them, and in my experience restricting too early can shut down things the model already does well without me asking.
A guide is a good starting point, but it isn't a rule. Copying it line by line is how you end up with a workflow that doesn't fit the way you actually work, and often burns more tokens than it needs to.
These are small changes and I'm still watching how they hold up, so the workflow keeps growing one piece at a time.