Context-Vigilance

Context Vigilance Essentials

1. What is a prompt?

• Role – who the AI should act as (e.g., “act as a product manager”).
• Goal – one thing you want done.
• Context – what matters (audience, tone, constraints, known facts).
• Output – the shape of the answer (bullets, table, plain text, JSON).

2. The shape of a clear prompt

1. One goal – don’t mix multiple tasks in one go.
2. Audience & tone – who will read it and how it should sound.
3. Boundaries – what to avoid; what to do when information is missing (e.g., “say ‘unknown’ rather than guessing”).
4. Output format – choose one: bullets, table, short paragraph, or simple JSON.
5. Example – give 1 short example of a good answer (and, if helpful, one bad example).

• Before: “Write about our product.”
• After: “Act as our product marketer. Goal: 5 bullet points for a landing page hero. Audience: startup founders. Tone: crisp and factual. Output: 5 bullets, each ≤14 words, including one risk/limitation. If something is unknown → write ‘unknown’.”

3. A simple prompt template (copy‑paste)

Role: {who should you be?}
Goal: {one clear task}
Audience & Tone: {who reads this, how it should sound}
Context: {facts, constraints, links if needed}
Output: {bullets | table | short paragraph | JSON}
Rules:
- If information is missing, say "unknown" (don’t invent).
- Keep it concise and concrete.
- Follow the Output exactly.
Example of a good answer:
{place a short example here}

4. How to write prompts (five simple rules)

• Start narrow. Ask for one thing at a time.
• Be concrete. Prefer numbers, ranges, and word limits over vague words like “detailed.”
• Say what not to do. Ban guessing, marketing fluff, or off‑topic content.
• Show the shape. Name the output and give a quick example.
• Keep it short. Short prompts are easier to maintain and improve.

5. How to improve quickly (lightweight loop)

1. Try on 3–5 real examples.
2. Mark what went wrong. Was it too long? Off‑tone? Missing structure?
3. Change one thing in the prompt (goal, tone, output, example).
4. Ask the AI to critique your prompt and propose two edits. Pick one and retry.
5. Save the better version with a clear name (e.g., “landing_hero_v2”).

• Fits the chosen output shape
• Concise and readable for the audience
• No invented facts when data is missing
• Meets the word/length limits
• Answers the single goal you set

6. Let the AI help you write prompts (meta‑prompts)

You design prompts. Based on the brief below, write a clear prompt using the template:
Role / Goal / Audience & Tone / Context / Output / Rules / Example.
Brief: {what I want}

Critique the prompt below: point out ambiguity and missing rules. Propose two improved versions and explain the difference.
<PROMPT>
{paste your prompt}
</PROMPT>

Generate 3 short example inputs and ideal outputs that match this prompt. Include one tricky case with missing information.

Rewrite the prompt so the output sounds {friendly | formal | neutral | bold} without adding fluff.

7. Output patterns (pick one and stick to it)

Output: 5 bullet points. Each ≤14 words. Include one risk.

Output: one paragraph (3–4 sentences). Avoid marketing language. Include one limitation.

Output: a 2‑column Markdown table with headers: Feature | Benefit. Max 6 rows.

Output: ONLY JSON: {"title": string, "audience": string, "risks": string[]}. If unknown → null.

8. Common pitfalls → quick fixes

• Too many goals at once → split into steps; run step by step.
• Vague words (“detailed”, “great”) → replace with counts, limits, or examples.
• No audience defined → name who will read it and adjust tone.
• No shape → state bullets/table/paragraph/JSON and give a 1‑line example.
• Guessing → explicitly say “if unknown, write ‘unknown’.”

9. Mini‑templates you can reuse

Role: analyst
Goal: Summarize the document for busy leaders.
Audience & Tone: execs, concise and neutral
Output: 5 bullets with facts (numbers, dates, risks). No marketing language.
Rules: If a fact is missing, write "unknown".

Role: support triage specialist
Goal: Assign a ticket to one of: [billing, bug, feature, account]
Audience & Tone: internal, factual
Output: "label: <one of the four>" and one‑line reason
Rules: If unclear, choose "account" and explain why.

Role: market researcher
Goal: Extract a brief snapshot about a competitor from the text
Audience & Tone: product team, neutral
Output: bullets: Company | Country | Products (max 3) | Pricing model
Rules: Do not guess; write "unknown" if not stated.
Example of a good answer:
- Company: Acme
- Country: Germany
- Products: mobile app; web suite
- Pricing model: freemium

Long Prompts & Example

Takeaway