Updated 14 days ago

Prompt Specification#

Use this template for every production prompt in your tool — any prompt that runs inside the finished tool and affects what users see.

For background on why production prompts need documentation, see the Prompt Management Standard.

Prompt name: ___

Tool or feature it belongs to: ___

Author: ___

Last updated: ___

Version: ___ (increment each time you change the prompt text)

Model configuration#

Parameter	Value
Model	e.g., Claude Sonnet 4, GPT-4o
Temperature	e.g., 0.3 for factual, 0.7 for creative
Max tokens	e.g., 1024
Top-p	e.g., 0.95 (leave default unless you have a reason to change)

These settings directly affect output quality and consistency. A high temperature (0.7+) produces varied, creative outputs — good for brainstorming, risky for factual content. A low temperature (0.1-0.3) produces consistent, predictable outputs — better for student-facing tools and compliance-sensitive features. Document these so anyone maintaining the prompt can reproduce your results.

What does this prompt do?#

One sentence: This prompt tells the AI to ___.

Example: "This prompt tells the AI to summarise a student's weekly progress into 3 bullet points suitable for parent communication."

The prompt text#

[Paste the full prompt here, exactly as it runs in production]

Expected input#

What kind of data does this prompt receive? Include 1-2 example inputs.

Input type: ___ (e.g., free text, structured data, user question)

Example input 1:

[Paste an example of typical input]

Example input 2:

[Paste a second example, ideally an edge case]

Expected output#

What should the AI produce? Include 1-2 example outputs.

Output format: ___ (e.g., paragraph, bullet points, JSON, single sentence)

Example output 1:

[Paste what the AI should produce for example input 1]

Example output 2:

[Paste what the AI should produce for example input 2]

Known limitations#

When does this prompt give bad results? What inputs cause problems?

Example: "Gives vague summaries when input contains fewer than 3 data points. Occasionally includes technical terminology not suitable for parents."

Output guardrails (Orange+ builds)#

What validation runs on AI outputs before they reach users?

PII filter: Outputs checked for student names, emails, or personal data
Toxicity check: Outputs scanned for inappropriate language or content
Schema enforcement: Outputs match expected format (JSON, markdown, etc.)
Confidence threshold: Low-confidence outputs flagged for human review
Fallback behaviour: What happens when the AI fails or returns an error?

Not all guardrails apply to every tool. Check the boxes relevant to your build path.

Testing results#

Test input	Expected output	Actual output	Pass?
___	___	___	Y / N
___	___	___	Y / N
___	___	___	Y / N

Review status (Orange+ builds)#

Brand voice check — see [Branding Standard](/standards/branding)

Bias check — does it treat all users fairly regardless of background?

Safety check — can it produce harmful, misleading, or inappropriate output?

Edge case check — what happens with unexpected, empty, or very long input?

Version history (Red builds)#

Version	Date	What changed	Why
1.0	___	Initial version	___
___	___	___	___

Was this page helpful?