🎯 Learning Objectives
- Control tone, voice, and writing style explicitly
- Manage output length with precision
- Produce structured outputs: tables, JSON, checklists
- Build guardrails that reduce hallucination risk
- Navigate the precision vs creativity trade-off
1. Tone, Voice, and Style Control
| Tone | Prompt Instruction | Use Case |
|---|---|---|
| Executive | "Direct, authoritative tone for a CEO" | Board presentations |
| Technical | "Precise technical language. Expert readers." | Regulatory docs |
| Conversational | "Explain as if to a smart friend, not a textbook" | Training materials |
| Persuasive | "Convince a sceptical audience. Lead with evidence." | Sales, proposals |
| Empathetic | "Acknowledge complexity. Avoid dismissive language." | Change management |
💡 Style Anchoring
Provide a sample: "Match the tone and structure of this example: [paste 2–3 sentences]." More reliable than describing tone abstractly.
2. Output Length Control
- Word count — "Respond in exactly 100 words"
- Sentence count — "Summarise in 3 sentences"
- Bullet count — "List exactly 5 risks" (forces prioritisation)
- Negative constraint — "Do not exceed 150 words"
⚠️ Always Quantify
Asking for "short" without a number produces inconsistent results. "Short" can mean 50 to 500 words depending on context.
3. Structured Outputs (JSON, Tables, Checklists)
JSON Output
Analyse the following SOP and return JSON with:
{"key_risks": [string], "compliance_gaps": [string],
"recommended_actions": [string], "severity": "high"|"medium"|"low"}
Return only valid JSON. No explanation text.
Table Output
Compare three validation approaches in a markdown table:
Approach | Cost | Risk Level | Timeline | Regulatory Acceptance
Checklist Output
Create a pre-audit checklist for a GMP inspection.
Format: numbered yes/no checkpoints.
Cover: documentation, equipment, training, and processes.
4. Guardrails Against Hallucination
- Uncertainty instruction — "If you are not certain, say so explicitly."
- Source flagging — "Where a claim needs a data point, note it as [VERIFY]."
- Scope restriction — "Only use information provided. Do not add external knowledge."
- Self-verification — "After writing, list any claims you are less than 90% confident about."
Important:
If you do not know the answer with confidence, say "I am not certain" rather than guessing.
Flag any regulatory figures or statistics with [NEEDS VERIFICATION].
5. Precision vs Creativity Trade-offs
| Need | Dial Towards | Techniques |
|---|---|---|
| Compliance documents | Maximum precision | Strict format, word limits, citation requirements |
| Creative campaigns | Maximum creativity | Open brief, multiple options, no format constraints |
| Executive communications | Precision-leaning | Clear structure, defined tone, specific audience |
| Brainstorming | Creativity-leaning | "Generate 10 different approaches to..." |
✏️ Module 04 Exercise
Take a recent AI output you were unhappy with. Diagnose it: Was the tone wrong? Too long? Unstructured? Did it hallucinate? Apply the specific guardrail technique for that failure mode and compare the outputs.
🔑 Key Takeaways
- Tone must be explicitly specified — the model's default is generic
- Always quantify length constraints — 'short' means nothing
- Structured formats turn AI into a usable, downstream tool
- Hallucination guardrails reduce risk but can't eliminate it
- The precision-creativity spectrum is a dial you control
- Self-verification prompts are one of the highest-leverage guardrail techniques