🏠 Home
Modules
📘 Module 01 📘 Module 02 📘 Module 03 📘 Module 04 📘 Module 05 📘 Module 06 📘 Module 07 📘 Module 08 📘 Module 09 📘 Module 10 📘 Module 11 📘 Module 12 ❓ FAQ 🚪 Sign Out
Module 04 of 12  ·  Hello Pharma AI Upskilling Program

Controlling Output Quality

Make outputs predictable, consistent, and usable.

📋 Official Content ⏱️ ~25 min read ✏️ Exercise included
📚 5 sections
🏢 Hello Pharma
🎯 5 objectives
⚕️

Hello Pharma AI Upskilling Program

This module is part of Hello Pharma's internal AI capability-building programme, designed to help every team member work with AI professionally and responsibly.

Official Content

🎯 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

TonePrompt InstructionUse 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

NeedDial TowardsTechniques
Compliance documentsMaximum precisionStrict format, word limits, citation requirements
Creative campaignsMaximum creativityOpen brief, multiple options, no format constraints
Executive communicationsPrecision-leaningClear structure, defined tone, specific audience
BrainstormingCreativity-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