orchestration · you, driving the agent
Day One
You can't one-shot this: the store's rules aren't in the brief, they're in the failures. The skill is the loop — run it, see the money that's wrong, tell the AI the rule, re-run, confirm. Someone who knows what to do reads the failures and closes it fast; someone who guesses ships wrong totals.
Your first day
real sandbox · the AI really rebuildsNo coding. You direct the AI: run the checkout, read which totals are wrong, tell it the rule it didn't know, run again. You win by getting every total right — and the game quietly watches whether you knew what to do.
step 1
Run it
The store's checkout already works — but its totals are wrong, and you don't know why yet. Run it to see what breaks.
step 2
Read the money
Each failure is in plain dollars: “shipping should be $6.00, your build gave $0.00.” That's a hidden rule you just discovered.
step 3
Tell the agent, re-run
Send the fix (click a suggestion or type it). The AI rebuilds. Re-run to confirm. Repeat until every total matches.
Why you can't just prompt it once:the store's rules — its free-shipping threshold, what it taxes — aren't in any brief. A one-shot prompt confidently ships the wrong totals because it doesn't have them. The only way through is the loop: run, see, steer, confirm. That loop is the skill.