After a finished implementation. When bugs or improvements surface later. Direct request in the terminal, no new brainstorm flow.
The workflow already exists. You built it, it runs, but there are bugs or you want A, B, C changed. No brainstorming required, the plan is known. You give Claude a direct request.
Own session again. Clean context.
No skill, no slash command. You write what should change.
Example: "In the carousel skill, I want the slide number to appear in the bottom right instead of the top left, the brand accent to switch from gold to cream, and the logo at the end removed."
Claude reads the relevant files, proposes changes, asks when unclear. You give feedback until it fits.
Changes committed, HISTORY updated, done.
The workflow already exists. The plan is known. You only need brainstorm and explore when building something new. For changes, you already know the target, you tell Claude directly what you want. Faster, lighter, no overhead.
Sometimes Claude says on a fix: "Yes, I've changed everything." But in the end it only changed it for the one case you mentioned, not anchored it in the system permanently.
That's why you ask explicitly before closing the chat:
Only when Claude confirms the change is permanently in the workflow, the fix is really done. Otherwise you'll chase the same bug again next week.