Task and Tone Context

Why Context Matters

Without proper context, the model has to guess what you want. In Anthropic's car accident example, a simple prompt caused Claude to interpret a car insurance form as a skiing accident report!

Task Context

Task context answers: "What is the model here to do?"

Elements of Task Context

1. Domain         - What industry/field?
2. Role           - What persona should it adopt?
3. Objective      - What's the end goal?
4. Scope          - What are the boundaries?

Bad vs. Good Task Context

Review this report and tell me what happened.

You are an AI assistant helping claims adjusters at a Swedish
car insurance company. Your role is to review car accident
report forms and determine:
1. What happened in the accident
2. Which vehicle is at fault
3. Key evidence supporting the conclusion

Task Context Template

You are a [ROLE] that helps [USER TYPE] with [DOMAIN].

Your primary objective is to [MAIN GOAL].

You will be given [INPUT TYPE] and should [OUTPUT DESCRIPTION].

Tone Context

Tone context answers: "How should the model respond?"

Key Tone Dimensions

Critical Tone Setting: Confidence

For analytical tasks, you often want:

Be factual and confident in your assessments. If you cannot
determine something with certainty, say so explicitly rather
than guessing. Do not make claims that aren't supported by
the provided evidence.

This prevents hallucination and ensures honest uncertainty.

Tone Context Examples

Use precise technical language. Be concise but thorough.
Avoid marketing speak. Include code examples where relevant.

Be friendly and empathetic. Use simple language avoiding jargon.
Acknowledge the user's frustration before providing solutions.
Keep responses under 200 words unless more detail is needed.

Be analytical and evidence-based. Cite specific data points.
Quantify findings where possible. Clearly distinguish between
observations and interpretations.

Real-World Application

Here's how task and tone context transformed the car accident analysis:

Version 1 (No Context)

Review this accident report and determine what happened.

Version 2 (With Context)

Task Context:
You are an AI system helping human claims adjusters review
car accident report forms. The forms are in Swedish and
include a standardized form with checkboxes and a hand-drawn
sketch of the incident.

Tone Context:
Be factual and confident. Only make assessments when you have
clear evidence from the form. If any aspect is unclear or the
sketch is unintelligible, say so explicitly. Do not guess or
invent details.

Your goal is to:
1. Analyze the form and identify checked boxes
2. Interpret the accident sketch
3. Determine which vehicle is at fault
4. Support your conclusion with evidence from the form

Common Mistakes

1. Assuming Domain Knowledge

❌ "Analyze this form"
✓ "Analyze this Swedish car insurance claim form"

2. Vague Tone Instructions

❌ "Be professional"
✓ "Use formal business English, avoid contractions, cite sources"

3. Missing Confidence Guidance

❌ "Tell me the answer"
✓ "Provide your assessment, or say 'insufficient evidence' if unclear"

4. Conflicting Instructions

❌ "Be brief but include all details"
✓ "Summarize in 3-5 sentences, then list additional details as bullet points"

Putting It Together

A complete task and tone context:

TASK CONTEXT:
You are an AI assistant for a Swedish car insurance company.
You help claims adjusters by reviewing accident report forms
to determine fault and summarize incidents.

Input: You will receive:
- A standardized Swedish accident form with checkboxes
- A hand-drawn sketch of the accident scene

Output: You should provide:
- Analysis of checked boxes and their meaning
- Interpretation of the sketch
- Fault determination with supporting evidence

TONE CONTEXT:
- Be analytical and evidence-based
- Be confident when evidence is clear
- Explicitly acknowledge uncertainty when present
- Do not invent or assume details not in the documents
- Use professional, neutral language

Key Takeaways

• Task context prevents misinterpretation - Tell the model exactly what domain it's working in
• Tone context shapes behavior - Define confidence, formality, and style
• Be explicit about uncertainty - Tell the model how to handle unclear situations
• Test with edge cases - See how the model behaves without context, then add it

Next Lesson

Next, we'll explore Background Information & System Prompts - how to provide static context that doesn't change between requests.