I’m looking for advice on how to prompt Gemini (or Gemma) to match the kind of responses I used to get from the original ChatGPT-4 model. As someone with dyslexia, I heavily relied on ChatGPT-4 for writing support. It was exceptionally good at catching and correcting subtle issues like:
* **Ambiguous pronoun references** (e.g., unclear “he,” “it,” or “they” usage)
* **Inconsistent verb tenses** (e.g., switching between past and present without context)
* **Homophone confusion** (e.g., mixing up “there,” “their,” and “they’re”)
Unfortunately, with ChatGPT-5 (or the more recent versions), I’ve noticed a decline in quality, especially in these nuanced areas. Previously, one-shot prompts worked well with GPT-4, but now I often have to retry multiple times or revert to tools like Grammarly.
Is there a specific prompting strategy I can use in Gemini to get responses similar to GPT-4’s earlier capabilities, particularly around nuanced language understanding and correction?
I’d love to hear from anyone experimenting with this, especially if Gemini can be tuned or prompted more effectively for this kind of language support.
Thank you for outlining this so clearly; it’s especially valuable to understand how critical nuanced language correction is for accessibility workflows. I really appreciate you raising this use case.
When working specifically with Gemma, consistency largely depends on how structured the instructions are. Because Gemma is an open-weights model, it responds best to very explicit task framing. A few strategies that tend to work well:
1. Define a strict rubric. Instead of a general “proofread this,” explicitly list the exact categories you want reviewed (example: ambiguous pronouns, tense inconsistencies, homophone misuse). Ask for structured output:
Quote the original phrase
Explain the issue
Provide a correction
This makes Gemma behave more like a precision editor rather than a general rewriter.
2. Separate detection from rewriting (two-pass approach).
First, ask Gemma to identify issues without changing your text.
Then, in a second prompt, ask it to produce a revised version incorporating only those corrections.
This helps prevent subtle issues from being silently rewritten and makes the feedback more transparent, which can be especially helpful when you’re relying on it for clarity checks.
3. Anchor with a short example. Including one brief example of the kind of correction you expect (even 2–3 lines) can significantly improve consistency.
4. Adjust decoding settings (if running locally or via API). For analytical editing tasks, lowering the temperature and using more deterministic sampling can improve reliability.
If this is a core workflow for you and you’re running Gemma locally, lightweight instruction tuning (example: LoRA adapters) on annotated editing examples can further improve performance, but structured prompting alone often gets most of the benefit.
Accessibility-centered applications like this are incredibly important, so thank you again for bringing it up.
thank you for the quick reply! I’m actually quite surprised! but this post was for the Gemini team! Apologies for posting wrong place. But good to see that Gemma team are paying attention to this space.