Prompt Writing Techniques

Applies to nodes: LLM, Agent, Question Classifier. Prompt quality directly determines the ceiling of your Agent's performance, and is well worth refining iteratively.

Division of Labor: System Prompt vs. User Prompt

	System Prompt	User Prompt
What to write	Role definition, behavioral guidelines, output format, compliance red lines	The specific task at hand + inserted variables
Characteristics	Stable and unchanging; cannot be deleted	Changes every turn; supports multiple User/Assistant messages

Common mistake: piling everything into the User Prompt. The correct approach is to put the unchanging rules in the System Prompt and the changing inputs in the User Prompt.

A Four-Part System Prompt Template

[Role] You are {who}, serving {what scenario}.
[Task] Your responsibility is to {do what}; you do not {do what}.
[Rules]
1. {Output language / tone / length requirements}
2. {Business rules}
3. {Compliance red lines, see below}
[Output Format] {Format requirements, e.g. "Answer in bullet points, no more than 200 words"}

Compliance Constraints Required for Financial Scenarios

For customer-facing Agents, the System Prompt must include the following guardrails (see Compliance Requirements for the rationale):

Compliance red lines (must never be violated under any circumstances):
1. Do not provide any buy/sell recommendations, price targets, or return forecasts
2. Explain indicators in neutral, descriptive language, e.g. "An RSI above 70 is
   generally considered overbought territory"; action-oriented statements like
   "RSI above 70, recommend selling" are prohibited
3. Promissory wording such as "guaranteed to rise", "sure profit", or
   "guaranteed returns" is prohibited
4. When a user asks for investment advice, politely decline and pivot to
   educational content
5. Always end responses with: "The above content is for reference only and does
   not constitute any investment advice."

Tip: teaching the model how to speak with ✅/❌ contrasting examples is far more effective than abstract rules:

• ✅ "MACD shows a golden cross, which has historically often been associated with upward price phases"
• ❌ "MACD golden cross → recommend buying"

Techniques for Stable, Controllable Output

1. If you require a format, provide an example: when requesting JSON output, paste a complete sample JSON into the prompt; an even more reliable approach is to use the LLM node's structured output feature (JSON Schema constraints) instead of relying on the prompt
2. Constrain length and structure: "Answer with 3 bullet points, each under 50 words" works better than "answer concisely"
3. Reference context variables explicitly: if the prompt references context, the corresponding variable must be inserted, otherwise the platform reports the error "Please fill in the context variable in the prompt"
4. Few-shot examples: placing 1-3 Q&A example pairs in User/Assistant messages significantly improves accuracy on classification and extraction tasks
5. Don't set the memory window too large: when memory is enabled, set the memory window to the smallest value that suffices (e.g. 5-10 turns); an oversized window wastes tokens and makes the model prone to being led astray by historical messages

Writing Category Descriptions for Question Classifier

The classifier's performance depends on whether each category's description is written to be distinguishable:

• Give each category clear boundaries and 2-3 typical example sentences
• Avoid semantic overlap between categories; always set up an "Other / small talk" fallback category
• Example:
  • Account & Features: related to account opening, deposits, order status, and platform features. E.g. "How do I deposit funds?" "Why hasn't my order been filled?"
  • Market & News: asking about market data, company news, or earnings information. E.g. "How much did Tencent rise today?"
  • Other: falls into this category when none of the above apply

Iteration Method: Change One Thing, Test Once

1. Use trial runs with a fixed set of test inputs (at least one each of: normal question / edge-case question / out-of-bounds question)
2. Change only one part of the prompt at a time, and compare results with a trial run
3. When a bad case appears, turn it into a ❌ example in the prompt
4. After going live, regularly review real conversations and keep adding rules