Directing AI, not using it
How I direct Claude Code agents and AI tools in two projects, with written rules, review gates and saying no, and what it means for leading a design team.
When I show work I made with Claude Code agents and AI tools, the claim is not "AI made it". The claim is: I set the rules, the approvals and the quality bar, and AI executes under them. I call that directing. It is a leadership skill: writing down who decides what, checking the work at set gates, and saying no.
My examples come from two projects. LOU is a narrative game I'm making alone. The PLL prototype started as my own pitch to PLL; now PLL's iOS and Android developers use its repo as the base for their production front-end code.
An agent does what the brief says, so the brief says who decides
On LOU, the brief names what only I decide: what becomes canon, when the mechanics change and whether a paid generation runs. The LOU case has the full rules, with the budget each map must fit.
The PLL prototype ran on one main rule: Figma leads, code follows. The cost is a Figma edit for every change, however small; the gain is a library and code that can't drift apart. Tests and lint hold the rule.
I check the work at gates, not by watching it
On the PLL prototype, every decision the agents take without me goes into a log with its reason, next to my own calls. A separate agent compared their work with my v2 Figma library, and new states came back marked as drafts for my review. The PLL case has the numbers.
On LOU, the gates are the automated tests, a balance simulator no agent may edit, a lore checker and a performance check for each map. I also play the game and look at every generated image before it goes in. One check, a lineup of the whole background cast, caught a fisherman built on a woman's body.
Gates have limits: on PLL, what the agents couldn't test, such as haptics and VoiceOver, needed me and a real iPhone.
The quality bar is what I say no to
Rules and gates catch errors. Whether the work is good enough is still my call.
On LOU, I rejected the first mood portraits on quality and paid more to make every mood again, rather than lower the bar.
On PLL, the agents built the full-width season banner the brief asked for, and it felt like an ad, so I replaced it. The agents did what they were asked, and the result still wasn't right.
Agent rules are decision rights, and I want them for a design team too
Each rule above is a decision right: what an agent may decide alone, and what comes to me with the evidence. On LOU, I approve or reject what an agent invents, and every session closes with a note for the next one.
I want the same with a design team: written decision rights, and handoffs that don't need a meeting.
That is why I call it directing, not using.