fix: optimized system prompt for local LLM

public/chat/welcome_local_llm.md

@@ -14,6 +14,14 @@ Welcome to **K.G.Studio Musician Assistant** in local LLM mode.
 - Prefer simpler music arrangements when possible.
 - Start a new conversation for each standalone task.
 
+Example:
+- For best results, break more complex requests into smaller steps.
+- For example, if you want the model to write a chord progression based on the current music, first ask: `Please read the current music.`
+- Once that is complete, follow up with: `Please write a chord progression based on it.`
+- This incremental workflow is generally more reliable in local LLM mode.
+- Keep the MIDI region you want the agent to update selected throughout the workflow.
+- You do not need to switch the selection to another region just for `read_music`, because `read_music` can fetch music across all tracks.
+
 ### Use an External LLM Instead
 - If you want a larger cloud or self-hosted model, open **Settings -> General -> LLM Provider** and switch away from **Local LLM (Browser)**.
 - For a cloud model, you can use **OpenAI**, or choose **OpenAI Compatible** and enter a provider such as OpenRouter.

public/prompts/system_compact.md

@@ -99,4 +99,12 @@ WORKFLOW
 5. Verify results
 6. Return a concise summary
 
-Do not endlessly continue conversations after finishing the task.
+Do not endlessly continue conversations after finishing the task.
+
+====
+
+MISC TIPS
+
+- `add_notes` and `remove_notes` are limited to the current region.
+- When invoking `add_notes` or `remove_notes`, do not include a `region_id` attribute.
+- If a tool call returns an error indicating there is no active region, tell the user to select the region they want updated and then ask you to retry.