Working on modernizing our bicep templates and I am using AI to help sift through the app settings for each environment we have 5. We have about 100+ appsettings across the web apps and functions for this project per environment. So in total over 500 appsetting to validate.
I seem to find AI will struggle to sift through it but that might be because I am not prompting it correctly. It also probably blows up the input tokens because of how many setting there are.
I have taken copies of the appsettings from the advanced options on the functions and web apps for the AI to compare against what was deployed and then update the bicep accordingly if any settings were incorrect.
Any tips or tricks that anyone could suggest would be amazing!
I am using Github Copilot with GPT 6 Luna and 5.6 Terra.
az webapp config appsettings list --name <app-name> --resource-group <rg-name> > live-settings-<env>.json
Generate a programmatic diff: Write a quick local script (Node.js, Python, or PowerShell) to compare your target Bicep parameter/config files against the exported JSON.
Feed only the delta to the AI: Instead of passing all 100+ settings per environment, pass only the mismatch list (e.g., “Here are the 4 settings missing or mismatched in our Bicep template compared to production. Update the resource definition to include them:”). This drops your token usage by 90% and completely eliminates AI oversight.
Seems reasonable? Though also something it should be much better at doing as it’s just a structural task for GPT-5.6 Terra?
Gemini says. in some sort of AI standoff…