Don't be blinded by LLMs - check everything #990
georgfranz
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
I use spec-kit and have tried it in PHP + Python projects.
The results are good.
But here's something I've stumbled across several times now:
You specify something, e.g., a statistics feature for a Python project. When specifying, you also use other LLMs to get a good basis and really good text.
Then you run /specify and follow the normal workflow.
The descriptions and texts are so good that you are completely dazzled by them, especially because they are very, very complex.
The problem: It has happened to me several times now that, as a programmer with almost 25 years of experience, I overlooked the fact that essential things were actually missing and not mentioned in these texts that were created. So they were not implemented either.
For example: I specified a statistics module to evaluate various logs in a very complex way. With API connection and all the trimmings. I was blown away by the documentation. Created with Sonnet 4.5 plan mode. Cross-checked with Codex and Gemini. Gemini 2.5 Pro wrote: He had never read such good planning before.
All three LLMs agreed that it was complete, so I started the implementation.
Then Sonnet said: Now it's feature complete and production ready.
Me: Are you sure? Check all the documents again. Yes, everything has been implemented. Codex asked: Yes, everything has been implemented. Gemini asked: That's really great!
Then I said to Sonnet: Then please try it out.
Then it turned out that it didn't work because the database connection had been forgotten.
He implemented all the queries—yes. But he simply didn't use the existing database connection for the project.
I hadn't mentioned this explicitly because I assumed that Sonnet was smart enough to check what was being used in the project and would automatically incorporate it. (Claude.md also states in abbreviated form: Don't reinvent what's already there, use libs, DRY + KISS).
It was as if Sonnet had built a house but forgotten the front door.
I don't know how to solve this either. You have to be even more attentive.
All in all, I can only give this tip: Don't be blinded by the excellent, very comprehensive texts; check everything thoroughly.
Beta Was this translation helpful? Give feedback.
All reactions