Two Demos, One Instance
I was building two customer demos at the same time, and both lived on one ServiceNow instance. That wasn’t laziness. It was the only instance I had with the full AI features enabled, and I wanted the AI beats in both demos.
Then one demo’s changes landed in the other demo’s update set.
Claude found it, not me. It turned up during a QA pass. Here’s what happened, the rules that came out of it, and what I still haven’t proven.
What happened
Two Claude sessions were building two different demos, each for a different customer, on the same instance. One session switched the model provider behind the AI features. Those changes landed as seven rows in the OTHER demo’s update set.
The cause was simple. Both sessions worked under the same admin account, and that account’s current application was the other demo’s app. The session wrote under whatever the account had selected last. It never set its own scope or update set first.
That isn’t really a prompt problem. The session did what the account pointed it at. The setup, one admin account, one current application, two builds in flight, is what allowed it.
The first call was to log the rows and leave them. A move attempt had failed and left an empty update set in the wrong scope, which was its own mess. Later that day, on my yes, one script moved all seven rows into sets for the right demo, and a read-back confirmed it.
The same session hit it again that day by a different route. The picker said Global, but the writes still landed in the other demo’s app. That’s why the first rule below says read it back.
The rules that came out of it
- Set your own scope and update set, and read them back, before ANY write. Use the demo’s app, or Global plus an update set named for the demo. State both in the report. Never write under whatever the shared account last had selected.
- Build one demo at a time, or give each demo its own build account. My instruction to Claude was: “we can’t build 2 demos in the same instance at the same time then, we need to single thread those to separate out the update sets, unless we create another account to run from, which is probably the ideal solution.” I create the accounts. A session never does.
- A row in the wrong update set gets logged, not moved. Moving or deleting it is a decision for me, not for the session that noticed it.
- One update set per demo build. ServiceNow holds one app scope per update set, so a build that touches several scopes gets one parent batch named for the demo, with one child set per scope. A platform change the build makes, like that model provider switch, goes in the batch too.
- Keep every build rebuildable. Shared instances get refreshed. Log every change and keep the update sets in order so the demo can be rebuilt.
- Inside an approved build, log it and keep going. I’m signed in as the admin, so Claude can do admin work as me, as long as every change is logged with how to undo it. It still asks first about anything outside the build, and about deleting records another person created.
- Email is never on.
That last one deserves a note. My rule: “email should never be turned on for any demo build, period.” There was no incident behind it. It’s a precaution, so we don’t accidentally spam people from a demo instance. Some rules don’t need a war story.
Shared means global
A shared instance has switches that change EVERY demo on it. Changing which model a capability uses by default needs a specific order: old default off, new default on, run one script, then read back that exactly one default exists. Turning off a built-in AI agent so a custom topic wins is instance-wide, so I time it around the other demos.
If one demo’s setup changes behavior another demo relies on, run the demos in date order.
Three gotchas from the build
- A 401 can’t tell a wrong password from a blocked method. A local build user with Basic auth on the REST Table API got a 401 on this instance, even with the right password. The SDK command line worked for the same user, because it signs in once and sends a session token with each call. We lost some time troubleshooting it before the method was suspected. Test with a second client that signs in differently before you reset anything.
- Never sign an SDK build account in to a browser. On this instance the MFA rules cover the admin role, so a web sign-in forced MFA enrollment, and the SDK then failed with a multi-factor error. Removing the account’s MFA records fixed it.
- The SDK’s add-auth command never checks the password. Two aliases were saved under the wrong user in one evening. Run one query per alias right after you add it.
What I haven’t proven
- Parallel builds on separate accounts. Each demo has its own build account now, but the builds have taken turns, not run together. An install as one user left a second admin’s current application and update set untouched, so I expect it to work. Separate accounts don’t fully isolate, though: a cross-scope privilege still landed in whatever update set a person had current.
- Rollback. Each install prints a rollback link. I haven’t run one.
- One writer per instance. Only one session writes to a shared instance at a time, and each hands the slot on when it finishes. It’s in use, but it runs on discipline, not a tool, so I can’t prove it’s never been bypassed.
The pattern
Every rule here came AFTER something broke, or after I worried it might. That’s IMO the right order. Don’t try to write the perfect list up front. Write down the thing that bit you, make it a rule, and put it where the next session will read it.
If you share a demo instance between builds, start with the first rule. Set your own scope, read it back, and make that a gate before any write.
Hope this helps!