This guide provides a structured approach to writing effective prompts for sheet AI actions. Well-crafted prompts lead to more accurate, consistent, and useful AI-generated outputs. Use simple prompts for simple tasks and more complex, structured prompts for complex tasks.
Key principle
Match prompt complexity to task complexity. A one-sentence instruction may suffice for simple tasks, while complex tasks benefit from structured prompts with detailed instructions, examples, and formatting specifications.
Prompt components
Every sheet AI action prompt should include relevant components from the following sections. Not all components are required for every task—use judgment based on complexity.
1. Background
Purpose: Define the context and purpose of the AI action. This helps the AI understand what role it's playing and what goal it should achieve.
What to include:
The overall objective or goal
Any relevant context about the domain, audience, or use case
The role the AI should assume (e.g., auditor, analyst, reviewer)
Simple example:
You are helping to identify inherent risks addressed by control activities.
Complex example:
Objective: As a SOC advisor, your objective is to perform the listed Test Procedures for the specified Control, utilizing all provided Documents as evidence. For each attribute, evaluate whether the requirements are met and, as a SOC auditor, provide an overall conclusion on whether the Control is satisfied based on the supporting documentation.
2. Instructions
Purpose: Clearly state what you want the AI to do. This is the most critical component of your prompt.
Best practices:
Be as specific and detailed as possible.
Break complex tasks into numbered steps.
For long input lists, clearly indicate start and end; instruct the AI to read the full list first.
Use consistent language throughout (e.g., always expect "yes" or "no" if asking for that type of response).
Specify any conditional logic or decision points.
Simple example:
In one to two sentences, write a summary of the inherent risks that the following control is helping to address.
Complex example:
Instructions: I am doing a mapping of SOC 2 Controls to a PCI DSS v4.0 Requirement. First, read the PCI DSS v4.0 Requirement below and each of the SOC 2 Controls. Then generate a response for which of the SOC 2 Controls are directly related to by having similar context to the PCI DSS v4.0 Requirements.
3. Response format
Purpose: Specify exactly how the output should be structured. This ensures consistency and makes outputs easier to use.
What to specify:
Length constraints (number of sentences, bullet points, paragraphs)
Structure (bullets, numbered lists, table, paragraphs)
What to include and what to exclude
Whether to include reasoning or just conclusions
Formatting preferences (avoid Markdown if needed)
Special symbols or indicators (✅, ❌, ⚪, etc.)
Simple example:
Provide your response in one to two sentences.
Complex example:
Response format: Only generate a response with the SOC 2 Controls that maps to the PCI DSS v4.0 Requirement. Include the full SOC 2 control description in the response. There may be more than one SOC 2 Control that maps to the PCI DSS v4.0 Requirement. Do not include any reason for the response and do not include the PCI DSS v4.0 Requirement wording in the response. If no SOC 2 Controls are directly related to by having similar context to the PCI DSS v4.0 Requirements, then just respond with 'No Mapping'.
4. Examples
Purpose: Show the AI exactly what you want by providing concrete examples. This is one of the most powerful ways to improve output quality.
Best practices:
Include actual data examples, not just placeholders.
Show both input and expected output.
Provide multiple examples if the task has variations.
Include edge cases if relevant.
Format examples exactly as you want the output formatted.
When to use examples:
When the desired output format is complex or specific.
When there are multiple acceptable formats and you need to show the preferred one.
When the task involves nuanced judgment calls.
When initial outputs aren't meeting expectations.
Example format:
Example Response Structure:
Control: [Control Name/ID]
Test Procedures Performed: [Brief summary of steps]
Attribute Assessment:
- Attribute A: [No exceptions noted ✅/Exceptions noted ❌/Not Evident ⚪] – [Concise supporting note]
- Attribute B: [No exceptions noted ✅/Exceptions noted ❌/Not Evident ⚪] – [Concise supporting note]
5. Inputs
Purpose: Reference the specific data sources and variables that the AI should use. This connects your prompt to the actual Fieldguide sheet data.
What to include:
Column references from your Fieldguide sheet (e.g., {Control Activity}, {Test Procedures})
Static text or data that should be included in every prompt
Related data sources or lookup tables
Any conditional inputs based on row data
Simple example:
Control: {Control Activity}
Complex example:
PCI DSS v4.0 Requirement: {Control Activity}
SOC 2 Controls: (Paste a unique list of all your client's secondary controls here)
Complete prompt template
Use this template structure to build your prompts. Remove sections that aren't needed for simple tasks.
[BACKGROUND]
Objective: [State the overall goal and context]
[Describe the role the AI should play]
[INSTRUCTIONS]
1. [First step or action]
2. [Second step or action]
3. [Continue as needed]
4. [Include conditional logic if applicable]
[RESPONSE FORMAT]
• [Specify structure: bullets, table, paragraphs, etc.]
• [Specify length constraints]
• [List what to include and exclude]
• [Specify formatting preferences]
[EXAMPLES] (Optional but recommended)
Example Input: [Show sample data]
Expected Output: [Show desired result]
[INPUTS]
[Field Name 1]: {Column Reference}
[Field Name 2]: {Column Reference}
[Static data or additional context]
Tips & tricks for better prompts
Start simple, then iterate
Begin with a basic prompt and test it. Add complexity only when needed. A well-crafted prompt may sometimes need refinement to get the best results. Don't try to create the perfect prompt on the first try.
Be explicit about edge cases
Tell the AI how to handle special situations:
What to do if data is missing or incomplete
How to handle null or empty values
What to do if no match is found (e.g., respond with "No Mapping")
How to handle ambiguous cases
Use consistent terminology
If you use specific terms in your instructions, use those same terms throughout the prompt. For example, if you say "assess whether the Attribute is met," don't later ask "evaluate if the requirement is satisfied." Consistency helps the AI understand exactly what you want.
Specify what NOT to include
Sometimes it's as important to tell the AI what to exclude as what to include:
Do not include any reasoning for the response.Do not include the original requirement wording in the response.Avoid Markdown formatting.Do not use symbols or special characters.
Test with real data
Always test your prompt with actual data from your Fieldguide sheet before deploying it across all rows. This helps you catch issues early and refine the prompt based on real results.
Control output length
Be specific about length to ensure outputs fit your needs:
In one to two sentences...Provide a concise summary of no more than 50 words.List three to five key points.Include only the control ID and description, nothing else.
Use professional language for professional tasks
For tasks like auditing, compliance review, or professional assessment, instruct the AI to use appropriate language:
Write as a SOC auditor, using professional, objective language.Use technical terminology appropriate for [domain].Focus on sufficiency and appropriateness of evidence.
Request structured decision logic
For tasks requiring judgment, guide the AI through the decision process:
First, read and understand X. Then, evaluate Y. Finally, determine Z.If condition A is met, respond with X. If condition B is met, respond with Y.Note any gaps or limitations if evidence is missing or incomplete.
Common pitfalls to avoid
❌ Vague instructions
Instead of: "Analyze this control"
Use: "Review the control activity and identify three key inherent risks it addresses, providing a brief explanation for each"
❌ Assuming AI knowledge
Don't assume the AI knows your specific processes, frameworks, or internal terminology. Provide context and definitions when needed.
❌ Overcomplicating simple tasks
Not every task needs a complex prompt. For straightforward tasks like "Summarize this in one sentence," keep it simple.
❌ Inconsistent formatting requirements
If your examples show bullet points but your instructions ask for a table, the AI will be confused. Make sure all components of your prompt align.
❌ Not testing before deployment
Always test prompts on a small sample before running them across hundreds of rows. This saves time and prevents having to redo work.
Recommended workflow
1. Define your goal
What specific output do you need from the AI? Be as precise as possible about what success looks like.
2. Assess complexity
Is this a simple task (one clear action) or complex (multiple steps, judgment required)? This determines how detailed your prompt should be.
3. Draft your prompt
Using this template, write out the relevant sections. For simple tasks, you may only need Background and Instructions. For complex tasks, use all five components.
4. Test on sample data
Run your prompt on 3-5 representative rows from your Fieldguide sheet. Check if the outputs match your expectations.
5. Refine based on results
Based on the test results, adjust your prompt:
Add examples if outputs are inconsistent.
Be more specific in instructions if outputs miss key requirements.
Clarify response format if structure is wrong.
Add edge case handling if unexpected inputs cause issues.
6. Deploy across all rows
Once satisfied with the results, run the AI action across your entire sheet. Spot-check a few additional rows to ensure quality remains consistent.
7. Document your prompt
Save your successful prompts for future use and reference. This builds a library of effective prompts for your organization. You can also save Sheet AI Prompts to your Engagement template so it will always copy into new engagements.
Final notes
Prompt writing is both an art and a science. The best way to improve is through experimentation and iteration. Don't be afraid to try different approaches, learn from what works, and refine your technique over time.
The goal is not to create the most elaborate prompt, but to create the most effective one for your specific task. Sometimes that's a single sentence; other times it's a detailed, multi-section instruction set. Match your effort to the complexity of the task at hand.
Good luck with your AI actions in Fieldguide! With well-crafted prompts, you'll unlock powerful automation and insights for your workflows.