Most of the posts on this site were drafted with AI assistance. I choose the topic, a pipeline I built turns the idea into a draft, and then I read it, correct it, and decide whether it goes up. I did it this way on purpose. I know what an LLM is good for, and I wanted to get ideas out of my head and into the world without writing every word by hand. An LLM is a tool like any other, and there is always a person in the middle.
Where it has gone wrong
Drafts get things wrong, and some of the mistakes were published before I caught them. In August 2026 I went back through the older posts and fixed what I found:
- One post built its whole approach on
@odata.oldValuefor SharePoint triggers. That is a Dataverse feature, and the SharePoint connector never sends it. I replaced it with a pattern that works: a sentinel column the flow sets after it finishes processing. - An offline-mode post passed an inline table to
SaveData, which needs a real collection, and its sync logic read a field that was never set, so it silently patched nothing. - A password post gave an example’s entropy as about 47 bits when its own assumptions work out to about 66.
- A flow-batching formula produced an extra, empty batch whenever the item count was an exact multiple of the batch size.
- Two claims of a 70% improvement had no source. I removed the number and described the mechanism instead.
What I do about it
I read every post before it goes up. Code samples and API calls get run or looked up, because a plausible function that doesn’t exist is the failure I trust least. Numbers I can’t source get cut, and so do anecdotes I can’t stand behind.
If you find a mistake, tell me.