Let AI Finish the Event Prototype Without Chasing Every Fix

You are an event promoter with a local show to launch. The flyer is ready, but the launch checklist and RSVP page still need work. You ask AI to build a prototype, a test version you can try before guests use it. It comes back looking good.

Then you add someone after the room is full, and the page gives them a confirmed spot anyway.

Now you are explaining the mistake, asking for a fix, and checking whether the fix broke something else. You wanted this job off your plate. Somehow you became the person reminding the assistant what the job was.

Hermes is an AI assistant that can use tools to do work. Its /goal command sets a finish line so it can keep checking and fixing within agreed limits. This fictional event is a private test on a computer, not a live signup service or customer result.

The first draft still owes you work

Say your brief is for Neighborhood Sessions: a local R&B night with room for 40 people. Supply the event details, launch tasks and made-up guest records. Give each guest a permanent ID to tell a repeat signup from a new person.

The first draft has the event details and tasks you can edit. But it confirms a guest after the room is full, then counts that guest twice. The page looks good and still gets the important part wrong.

Hermes keeps the same goal as it works. After each turn, or round of work, another AI checks the last response against that goal. This second AI is called the judge.

If the job is still unfinished, it can send Hermes back to work in the same chat. The loop stops when done, pauses when blocked or out of turns, or waits when you pause it.[1]

Hermes needs to fix the waitlist, count each guest once, then check again. You should not have to type “keep going” to get a requirement you already gave it. Bring back the page, test results and questions that need your call.

Decide what would convince you

“Make an event app” leaves the assistant guessing. You decide what the page must do and what would convince you it works. Write that down with the limits and reasons to stop. Hermes calls this a completion contract. I would start with this brief.

Hand this test assignment to the person setting up Hermes. It is not a ready-to-run service. Review it and set the controls before running /goal, which starts work right away.

Outcome: build a local Neighborhood Sessions test page from the made-up event details and guest records only. Show event details, launch tasks I can edit, confirmed guests and a waitlist. Label every screen as a prototype. No real guest information.

Verification: fill the room, then add one more guest and show them on the waitlist. Add the same event-and-guest ID twice and prove it counts once. Edit a launch task, reload the page, and show the edit is still there. Try the controls with a keyboard and on an upright phone screen. Return the test page and results for each check. Say what you did not test.

Boundaries: change files only in the separate test folder. No public signup, messages, bookings, payments, pricing changes or publishing. Stop if event details are missing, capacity figures conflict, you need access outside that folder, or you reach a limit. Bring me the question instead of making up a venue rule.

The person setting up Hermes must limit which files it can change and which tools it can use. They must remove access keys for live services and turn off outside actions, such as sending messages.

For this test, use a four-turn goal budget and a $2 total cap for both the assistant and judge. The billing or execution system must enforce that cap. These are example limits, not a price quote. If the cap cannot be enforced, do not start the paid loop.

A turn budget is not a dollar limit. And writing “do not publish” does not remove a publishing tool. The setup must enforce those limits outside the prompt.

Hermes supports registered quality gates: code checks that must pass before it marks the goal done. Asking for a capacity test does not register one; the person doing setup must add it. A passing check only proves what it tested.[2]

The program running Hermes must remain alive while it works. A saved goal lets you pick up where you left off. It does not keep working while the computer is asleep or the program is stopped.

Why I use goals for builds

I used Hermes /goal for the initial build of the live cYpher.city Arcade on cYpher.camp. I would put it at roughly 80% of the way there. That is my estimate, not a count of the code. I still had to review the work, connect the pieces and improve it.

That is why I care about getting past the first good-looking screen. The Arcade has four learning districts: AI Heights, Market Street, Firewall Quarter and Chain Docks, leading to Agent Tower. Its reward campaign is currently unavailable; badges and XP are not AI credits. The link shows the game after later work, not independent proof of what happened in my first run.[3]

The Arcade brief, reconstructed

This is a reconstructed brief, not my original prompt or a record of bugs from that run: make each district reachable, keep each tower’s learning requirements, make battles work by touch on an upright phone, and keep saved progress working after reload. Return the local game and check results. Do not publish, connect accounts or promise rewards.

Your event page needs different checks: capacity, repeat signups and saved tasks. The point is the same. Give enough direction to finish without making you find every missing piece.

Watch the work continue

The event-launch script below follows that repair. The other tabs cover common event questions and sponsor inquiries. Both show when to stop and ask you. All three are fictional, with full scripts you can read without playing the animation.

You accept the work, then delegate the routine

The judge can stop too early or keep going after the useful work is done. You accept the work. Try the page and inspect the test results first; a confident summary cannot tell you whether the waitlist works.[4]

But I would not have you approve the same capacity check forever. Once the checks and controls work, give permission for that routine in advance. After an edit, Hermes can rerun the checks and fix failures within the same limits.

Keep a record and review samples now and then. Test that problems reach you and that you can pause the loop. Pausing stops further rounds; it does not cancel an action already underway.

Changing capacity, making up a venue rule, merging sponsor records or opening public signup still needs your separate decision. Passing the tests grants none of that permission.

Judge the trial by how much chasing remains, including your review time and AI spending. If you still direct every repair, narrow the job or improve the checks. The event still needs your judgment. Repeating instructions should take less of your day.