# User: any idea why, all of a sudden, superwhisper has started acting up and not simply transcribi... ## Retrieval Notes - Session id: `613258bc-da26-47a3-80ac-fdfdced0324d` - Last activity: `2026-01-27T20:46:17.907Z` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User User: any idea why, all of a sudden, superwhisper has started acting up and not simply transcribing my speaking but answering like a noral llm would? I didn't change the prompt (here it is anyway) and I didn't change the transcribing module Parakeet, nor the underlying llm sonnet 4.5 Assistant: Let me read your SuperWhisper prompt to see what you're working with.Your prompt looks bulletproof — explicitly forbids conversational responses with examples. So the issue isn't the prompt content. A few likely culprits: **1. SuperWhisper app update** Check if it auto-updated recently. App updates sometimes reset settings or introduce new "modes." Look in SuperWhisper's settings for: - A mode toggle (transcription-only vs. AI assistant mode) - A [Thinking: 2 block(s), 10.0s total] [Tool Read input: file_path=/Users/nickyoung/My Obsidian Vault/_Legacy/Personal/Backups/superwhisper prompt.md status=completed] User: you got cut off, please continue from where you left off. ### Assistant Right, continuing from where I was: **1. SuperWhisper app update** Check if it auto-updated recently. App updates sometimes reset settings or introduce new "modes." Look in SuperWhisper's settings for: - A mode toggle (transcription-only vs. AI assistant mode) - A "post-processing" or "AI enhancement" toggle that might have been enabled - Any new "smart" features that got turned on by default **2. Anthropic API-side changes** Sonnet 4.5 might have received updates that make it more "eager to help." Claude models sometimes get tuned to be more conversational. Your prompt is strong, but if the model's baseline behavior shifted, it might be overriding instructions more readily. **3. Temperature or sampling settings** Did SuperWhisper's settings panel add new options? A higher temperature could make the model more likely to deviate from strict instruction-following. **4. Context window pollution** If SuperWhisper maintains any conversation history or context between transcriptions, old assistant-style responses could be leaking into the context and priming the model to respond similarly. **Diagnostic steps:** 1. Check SuperWhisper's version history / changelog 2. Try switching to a different Claude model temporarily (Haiku?) to see if the behavior persists 3. Look for any "AI mode" toggle in settings 4. Check if there's a "clear context" or "reset" option **Quick test:** Add this to the very start of your prompt (before everything else): ``` CRITICAL: Output ONLY the corrected transcription. No preamble, no explanation, no conversation. Just the text. ``` If that doesn't work, the issue is almost certainly app-side, not prompt-side. *Se il prompt non è cambiato e nemmeno il modello, il colpevole è probabilmente un aggiornamento silenzioso dell'app.*