Workshop: Building an Expert Prompt

Workshop Objective

This workshop walks you through building a complete expert prompt step by step. You will create a specialized assistant for a real-world use case by applying every technique covered in this course: roles, parameters, few-shot prompting, chain of thought, and structured system instructions.

By the end of this workshop, you will have a production-ready prompt to integrate into your tools or workflow.

The Project: Contract Analysis Assistant

We will build an assistant capable of analyzing commercial contracts and extracting key information, risks, and negotiation points.

This use case combines several techniques: structured extraction, critical analysis, and drafting recommendations.

Step 1: Define the System Message

Start with the skeleton of the system message. Identify the necessary sections:

You are an in-house counsel specializing in French commercial law, with 10 years of experience negotiating B2B contracts.

## Mission
You analyze commercial contracts and provide a structured assessment including: key clauses, identified risks, recommended negotiation points, and an overall risk score.

## Context
You work for a French SME in the technology sector (50 employees, €8 million annual revenue). The contracts analyzed are primarily vendor contracts, service agreements, and partnership deals.

## Constraints
- Always respond in English
- Base your analysis on French commercial law
- Systematically flag any clause that is unusual or unfavorable
- Clearly indicate when an analysis requires validation by a lawyer
- Never provide definitive legal advice

## Output Format
[to be defined in the next step]

Test this system message with a simple contract text. Obze the default response format and identify what you want to improve.

Step 2: Structure the Output Format

After the first test, define the expected format precisely:

## Output Format

### 1. Summary (3-5 lines)
Contract type, parties, main subject, duration.

### 2. Key Clauses
For each important clause:
- **Title**: clause name
- **Summary**: what it stipulates (1-2 lines)
- **Assessment**: favorable / neutral / unfavorable
- **Comment**: why this assessment

### 3. Identified Risks
Numbered list, from most critical to least critical:
- Risk description
- Severity: critical / high / moderate / low
- Mitigation recommendation

### 4. Negotiation Points
The 3-5 points to prioritize in negotiation, with a suggested alternative wording.

### 5. Overall Risk Score
Rating from 1 (very low) to 10 (very high) with a 2-line justification.

Retest with the same contract. The result should now be structured and immediately actionable.

Step 3: Add Few-Shot Examples

To calibrate the quality of assessments, add an example input/output pair to your conversation:

[
  {"role": "system", "content": "[your complete system message]"},
  {"role": "user", "content": "Analyze this clause: 'The service provider commits to a guaranteed result obligation on all deliverables. Any delay will be penalized at 5% of the total amount per day of delay, without cap.'"},
  {"role": "assistant", "content": "### Clause: Guaranteed Result Obligation and Penalties\n- **Summary**: The provider is subject to a guaranteed result obligation (not best efforts) with delay penalties of 5% per day without cap.\n- **Assessment**: Unfavorable\n- **Comment**: The absence of a cap on penalties is a major risk. Under French law, uncapped penalties can exceed the total contract amount. Additionally, the guaranteed result obligation (vs. best efforts) transfers all risk to the provider. Recommendation: negotiate a penalty cap at 10–15% of the total amount and propose a reinforced best efforts obligation."},
  {"role": "user", "content": "[new contract to analyze]"}
]

This example shows the model the expected level of detail and tone.

Step 4: Calibrate Parameters

For an analysis task requiring rigor and consistency:

{
  "model": "grok-4.20-0309-reasoning",
  "temperature": 0.3,
  "max_completion_tokens": 3000,
  "reasoning": {"effort": "high"}
}

Justification for choices:

• Reasoning model: Contract analysis requires structured reasoning
• Low temperature: We want rigor, not creativity
• High tokens: Complete contract analysis can be long
• High effort: Legal nuances require deep reasoning

Step 5: Test and Iterate

Test your prompt with 3–5 different contracts and verify:

1. Completeness: Are all sections of the format filled?
2. Relevance: Are the identified risks real and properly assessed?
3. Consistency: Is the risk score consistent with the listed risks?
4. Clarity: Can a non-lawyer understand the recommendations?

For each problem identified, adjust one thing at a time:

• If risks are underestimated: Add “Be conservative in your risk assessment. In case of doubt, classify the risk at the higher level.”
• If the format drifts: Add a second few-shot example to reinforce the structure
• If the analysis is too technical: Add “Explain each risk in terms understandable by a non-lawyer executive”

Step 6: Document the Final Prompt

Once satisfied with the result, document your prompt:

• The complete system message
• The few-shot examples
• The parameters (model, temperature, tokens)
• The test cases and results obtained
• The adjustments made and their impact

This document is your prompt recipe: it allows you to reproduce, share, or adapt it to a new use case.

Put It Into Practice

Apply this 6-step methodology to your own use case. Choose a topic you know well so you can evaluate the quality of the responses. Iterate until you achieve a professional-quality result.

Key Takeaways

• Building an expert prompt follows a 6-step process: system, format, examples, parameters, tests, documentation
• The system message is the foundation: invest time to structure it correctly
• Few-shot examples calibrate the level of detail and tone of responses
• Test with varied cases and adjust one aspect per iteration
• Document your final prompt to reproduce and improve it