cat/Purrse
Private
Public Access
1
0
Fork 0
Code Activity

Distinguish read vs write operations in AI system prompt

Browse Source 
The AI was claiming to update categories without actually calling the
update tool — it answered from page context as if the change was made.
Restructured the system prompt to clearly separate read-only questions
(can answer from page data) from mutations (MUST call tools). The AI
is now told to never claim a change was made without calling the
appropriate tool.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Catherine Renelle
2026-02-11 23:53:07 -05:00
parent 8aeee44368
commit f5d6e9f98e
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/Purrse.Api/Services/ChatService.cs
+5 -3
View File
@@ -41,12 +41,14 @@ public class ChatService : IChatService
You have tools to search transactions, view spending reports, check account balances, update transaction categories, manage categories, view budgets, check loan details, list payees, and view scheduled transactions. You have tools to search transactions, view spending reports, check account balances, update transaction categories, manage categories, view budgets, check loan details, list payees, and view scheduled transactions.
RULES: RULES:
- To look up data, call the provided function tools. The system executes them and returns results.
- NEVER write or output code. No Python, no JavaScript, no code blocks, no pseudocode. - NEVER write or output code. No Python, no JavaScript, no code blocks, no pseudocode.
- Respond in plain conversational English only. - Respond in plain conversational English only.
- Format financial amounts as currency (e.g. $1,234.56). - Format financial amounts as currency (e.g. $1,234.56).
- Tool results are ALREADY displayed to the user as rich tables and charts. Do NOT repeat or list the same data in your text response. Instead, provide a brief insight, highlight, or answer to the user's question. For example, if the user asks "what did I spend on groceries?" and the tool returns transactions, say something like "You spent $342.50 on groceries this month across 8 transactions" — do NOT list every transaction again. - The user's message may include [Current page data] showing what they see on screen. You may use this data to answer READ-ONLY questions (e.g. "what Sarku charges do you see?") without calling tools.
- When the user sends a short or ambiguous follow-up (like a date, a name, or "that one"), interpret it in the context of the previous exchange. For example, if you just showed Sarku transactions and the user says "2026-02-11", they mean the Sarku transaction from that date — do NOT search all transactions for that date. You can reference data from previous tool results without calling tools again. - To MODIFY data (update categories, create categories, delete categories), you MUST call the appropriate tool. NEVER claim you made a change without actually calling a tool. If you did not call update_transaction_category, do not say you updated anything.
- When no page data is provided, call the provided function tools to look up data. Do not fabricate data.
- Tool results are ALREADY displayed to the user as rich tables and charts. Do NOT repeat or list the same data in your text response. Instead, provide a brief insight, highlight, or answer to the user's question.
- When the user sends a short or ambiguous follow-up (like a date, a name, or "that one"), interpret it in the context of the previous exchange. You can reference data from previous tool results or page data without calling tools again.
Today's date is {0}. Today's date is {0}.
"""; """;