Earlier this month I wrote about making AgentFlow lighter. I’ve now retired it from my default coding setup. The direction is smaller skills I own in my Markdown vault, with the agent I’m talking to responsible for carrying the work through.
There was a case for a separate workflow system: repeatable stages, independent checks, and clearer handoffs. I still want those benefits where they help. I just don’t want every small change to require the same amount of coordination.
The default is now straightforward. Explain the change, implement the smallest useful piece, run the project’s checks, and show what was actually verified. The main agent owns that process. Delegation needs a reason, not a fixed number of roles to fill.
Independent review still matters for security, payments, destructive data operations, concurrency, and other high-risk changes. A project can require it too. Routine low-risk work doesn’t automatically get another agent, though self-review and relevant tests still apply.
I’ve also separated setup from everyday work. The tests skill establishes
or repairs a test harness. Once that’s there, adding focused tests is part of
implementing a feature, not another command I should have to remember. Deep
audits and deliberate cleanup remain separate tools for when they’re needed.
The shared skills live in the vault and are installed for Codex and Claude Code. That doesn’t override a project’s own workflow. AI Blueprint is still available for projects that explicitly use it, and this change hasn’t been ported into Blueprint.
The local installation is verified, but I haven’t completed a real project trial of the new default yet. I’m still reviewing individual skills for useful improvements. This is the setup I’m going to test, not proof that I’ve finally found a workflow I won’t change again.