Prompting Fundamentals
Why Prompts Matter
Prompts are the core instruction set that tells the AI model how to act.
A well-structured prompt ensures accurate, consistent, and efficient results.
In V7 Go, every property that uses an AI Model needs a prompt.
This is where you define the logic, formatting, and nuance of the extraction.
Prompts allow you to have the greatest amount of customization, and these are typically where our solutions team spends the most time tweaking workflows to get the best results.
Prompting Fundamentals
Prompt Structure: Role → Task → Format
This is the standard and most effective structure used across Go:
| Component | Purpose | Example |
|---|---|---|
| Role | Tell the model who it is | “You are an expert legal analyst.” |
| Task | Define what the model should do | “Extract the lease start and end date from the document.” |
| Format | Specify exactly how the result should appear | “Return a JSON object with keys: start_date, end_date.” |
Example Prompt
Prompt Structure: Reasoning
This prompt structure is designed for properties that need reasoning steps built in.
Component | Purpose | Example |
|---|---|---|
Role | Tell the model who it is | You are a senior tax compliance analyst specializing in corporate reporting and regulatory audit preparation. |
Objective | Tell the model what it’s trying to do. | Your task is to extract the following compliance-critical data points from the uploaded document:
|
Context | Provide the prompt context around what you’re trying to solve. | Steps |
Steps | Specify exactly how the LLM should think through this problem, step by step. | First, identify and isolate the relevant sections of the document where each value is discussed or reported. Next, for each field:
|
Output Format | Specify exactly how the result should appear | Return your result as a JSON object using the following format: { |
Example Prompt
Prompt Structure: Advanced Reasoning
This prompt structure is designed for properties that need advanced reasoning steps that require specific success criteria and areas to stay away from:
Component | Purpose | Example |
|---|---|---|
Role | Tell the model who it is | You are an experienced investment analyst at a private equity firm, trained to evaluate CIMs and summarize actionable insights for the deal team. |
Objective | Tell the model what it’s trying to do. | Your task is to extract and synthesize the most critical information from the attached CIM and generate a memo that covers investment thesis, financial highlights, risk factors, and potential red flags. |
Context | Provide the prompt context around what you’re trying to solve. | The CIM contains detailed business descriptions, market analyses, financial projections, and risk disclosures. Your memo will be reviewed by internal stakeholders to decide whether to advance the opportunity to diligence. It must be high-signal, concise, and decision-oriented. |
Steps | Specify exactly how the LLM should think through this problem, step by step. |
|
Output Format | Specify exactly how the result should appear | Return your answer as a structured text memo with the following sections: CIM Summary Memo
|
Guardrails | Tell the model what it should avoid, or not reference/produce in it’s output. | Guardrails
|
Success Criteria | Specify what a successful output should include, including any critical information or must-haves. | Success Criteria
|
Example Prompt
Prompting for JSON Properties
For JSON properties, be explicit about the expected schema:
Prompt:
This improves:
- Accuracy: LLM knows what to look for
- Consistency: Output can be reliably parsed
- Efficiency: Reduces need for multiple calls
Top Tips
- Avoid prompt ambiguity: Be precise about field names and expected formats
- Bundle similar values in one prompt when using JSON (improves cost + accuracy)
- Reference token-saving tactics: Less back-and-forth = lower cost
- Use examples in prompts when you’re seeing hallucination or inconsistency
Key Takeaway
Strong prompts = stronger workflows.
They reduce hallucination, improve citation grounding, and drive reliable outputs at scale.
Always:
- Define the role
- Be specific in the task
- Lock down the output format
Use this as your starting point for developing a standard structure in Go:
You are a [ROLE].
[DO THIS TASK].
Return the result in [FORMAT].
Updated 8 days ago
