I don't start with a roadmap. I start by watching someone actually stuck — my wife re-keying figures into a spreading template by hand, working out a covenant ratio against a deadline she didn't set. Or me, rereading an email because the tracker no longer matched what was actually happening. If I can't point to a specific moment someone hit a wall, I don't trust the feature enough to build it.

None of this is a separate skill I picked up because AI is trendy, either. Writing acceptance criteria, deciding what "done" actually means, and reviewing work before it ships is what I've done for a living for years — directing Claude Code is the same discipline, pointed at a different kind of team member. The bar I'd hold a sprint deliverable to is the bar I hold an AI-generated change to. If something comes back marked "fixed," I ask what I'd ask any developer: fixed where, exactly, and is that the part that actually matters or just the part that was easy to check. Deciding what ships this week versus what gets flagged as a deliberate deferral is scope management I'd run on any programme, payments or otherwise.

That same instinct decides how much process I put around a piece of work in the first place — I go deeper on the two very different governance models I actually run on my Delivery Approach page.