The Issue
Last week, I tried to use Perplexity AI to teach me about EAS Builds. I got thrown off everywhere was confused for a long time! Here're the sequence of questions I asked:
Then
It's all well and good, within the context of EAS Build variations.
Problematic follow-up question
I assumed the below is also within the context of the conversations above, but it gave me a surprising answer:
And the first reference returned was a 2019 article talked about unrelated structure - preview, production and testing branching of the source code on GitHub, not EAS Build config.
What I thought I asked
- I’m talking about
eas.json - EAS build profiles (
development,preview,production) -
distribution: "store" - TestFlight vs App Store release
This is a binary signing & distribution question.
What Perplexity answered (what it assumed)
Perplexity cited this article:
How to manage staging and production environments in a React Native app
At first glance, it sounds relevant:
- staging vs production
- React Native
- real-world apps
But that article is about:
- environment variables
- API base URLs
- runtime configuration
- feature flags
👉 It has nothing to do with EAS Build, TestFlight, App Store Connect, or signing.
Why Perplexity didn’t “infer” my earlier context
This was the key realization:
Perplexity does not reliably carry context across separate queries.
Each question is treated as a fresh retrieval task:
- interpret keywords
- fetch high-ranking related pages
- synthesize an answer
If most indexed content around “staging vs production” is about environment variables, that’s what it will pull — even if you meant EAS build profiles.
The fix: anchoring the context (aka context engineering)
What actually works is pinning the domain so the AI can’t drift.
Things that help a lot:
- naming concrete artifacts:
eas.json - naming commands:
eas build,eas submit - naming platforms: TestFlight, App Store Connect
- explicitly excluding irrelevant meanings
Example of a much better prompt:
I am using Expo EAS. In
eas.json, I have build profiles likedevelopment,preview, andproduction.
Explain the difference between a preview build and a production build in terms of signing and distribution (TestFlight vs App Store).
Ignore articles about staging/production environment variables.
Even better: paste a snippet of the eas.json.
Conclusion
There is still some technical details to learn about AI tools, even though it uses plain English and feels "easy" to use! I definitely need to learn more for it to be useful in solving my everyday issues.
Top comments (2)
This is such a relatable AI moment 😄 I’ve had similar issues where the tool answered the keywords I used, not the actual problem I meant. The idea of “anchoring context” with concrete terms like
eas.json, commands, and platforms is a really practical takeaway. Great reminder that with AI, how we ask is often more important than what we ask.Yes I now realised that context is super important! Good thing that Perplexity gave an original link reference, otherwise I would have gotten the wrong information.
It seems also that a lot of time the information could be from outdated documentation. So definitely a link to the current SDK version in the prompt would be super helpful.