My AI workflow is pretty simple, but it's also very powerful and effective. It contains only the following parts:
Start with the model + harness
Pick the frontier model provider + harness of your choice. I switch between them as need/interest dictates, and this setup works with all of them, including having the mid-tier models ship entire features or collections of features.
Just add these skills
You don't need a sprawling collection of plugins or heavyweight orchestrators. Five atomic skills provide everything needed for end-to-end development:
goal
Persistence
Does what you ask, no matter how long it takes, until it's done.
clarify
Alignment
Like grill-me, but even simpler, as it trusts the smarter models we have today to work out many of the specifics instead of me needing to answer more questions.
rebuild
Verification
Detect and run test suites/ci, linting tools, binary builds (if applicable).
wrapup
Release Engineering
Adds any changed files to source control, reviews .gitignore in case it needs to be updated to ignore any new sensitive (credentials) or ephemeral (logs, build artifacts), bumps the version, and commits + pushes code.
redeploy
Production
Idempotently deploy or redeploy the application I'm working on. I usually create a bin/deploy script that's specific to the application so that the model just has to figure out that it just needs to call that script, but some applications are more complex.