You fixed the same phrasing three jobs in a row, and on the fourth job the AI wrote it wrong again. It feels like the tool ignored you. It didn't. It never remembered the first three corrections in the first place, and once you understand why, the whole "the AI is stubborn" feeling turns into something you can actually engineer around.
Why does a language model forget your corrections?
A trained model's knowledge lives in fixed weights that do not change when you chat with it. The weights were set during training and then frozen; a normal conversation does not update them (Anthropic, Prompt engineering overview). So when you correct an output and start a fresh job, that correction is gone unless something deliberately carries it forward. The model is not being lazy. It has no memory of your edit because nothing wrote your edit anywhere it can see.
The thing people picture, "the AI learned from my correction," is really model retraining, and retraining a large model is expensive, slow, and not something that happens between two of your inspections.
So how do you make an AI actually use what you taught it?
You feed the lesson back in at the moment it writes, as an example. This is called in-context learning: a model adjusts its output based on examples placed in the prompt, without any change to its weights (Google Cloud, What is in-context learning?). Anthropic's own guidance is blunt about how strong this is. Showing the model three to five examples of the output you want, called multishot or few-shot prompting, is one of the most effective ways to get the format and voice you are after (Anthropic, Multishot prompting).
Read that twice, because it is the whole trick. The model did not "learn" your preference in any permanent sense. It matched a pattern you put in front of it. If your correction from last month is sitting in this month's prompt as a worked example, the model follows it. If it isn't, the model has no way to know it ever existed.
Why this matters more for a mold report than for a chatbot
A mold inspector's corrections are not stylistic whims. They are how you phrase a Category 3 finding so it survives a claim, how you word a limitation so it does not over-promise, the way you describe an indoor:outdoor comparison so an adjuster reads it correctly. Those are reusable decisions. If every report starts the model from zero, you re-make and re-fix the same decisions on every single job, which is exactly the grind that makes AI feel like more work than writing it yourself.
The inspectors who get value out of AI are not the ones with a magic prompt. They are the ones whose past corrections keep showing up in the next draft.
Capture each correction and replay it
The fix is to capture each correction as an example and replay it automatically. When you edit a draft, the change becomes a worked before-and-after pair. On your next report, the relevant pairs go back into the prompt as few-shot examples, so the model starts from your corrected version instead of its generic default. Nothing gets retrained. The model just keeps seeing what you already decided.
This is the design behind MoldMind's correction learning. Your edits are tracked per account and fed back as examples on future jobs, so a fix you made once does not have to be made again. Because the standards layer stays stable across jobs, that part of the prompt can be cached, which keeps the cost of carrying your history low (Anthropic, Prompt caching). The point is not that the AI replaces your judgment. The point is that your judgment stops evaporating after every report. You can see the kind of output this produces in the sample report.
Sources
- Anthropic, Prompt engineering overview: a trained model's weights are fixed; ordinary prompting does not update them.
- Anthropic, Multishot prompting: three to five in-prompt examples are among the most effective ways to control output format and voice.
- Google Cloud, What is in-context learning?: models adapt output from in-prompt examples without changing their weights.
- Anthropic, Prompt caching: a stable prompt prefix can be cached to reduce the cost of carrying repeated context.