Universal Prompt Patterns: Rewrite, Summarize, Extract, Classify (12 Templates)

Quick take: These universal prompt patterns let teams move fast without chaos. You get 12 templates across rewrite, summarize, extract, and classify, each with JSON outputs, guardrails, and role or audience switches. Copy, adapt, and standardize.

Base JSON schemas

Use these minimal schemas across patterns to keep downstream automation stable.

Rewrite schema

{
  "text": "string",
  "notes": "string"
}

Summary schema

{
  "bullets": ["string"],
  "actions": [{"owner":"string","task":"string","due":"YYYY-MM-DD"}],
  "risks": ["string"]
}

Extraction schema

{
  "entities": [{"type":"string","value":"string"}],
  "fields": {"customer":"string","product":"string","amount":"number"},
  "confidence": 0.0
}

Classification schema

{
  "label": "string",
  "confidence": 0.0,
  "reason": "string"
}

Global guardrails to include in every prompt

• No new claims. If numbers appear, add a short “Sources:” note or return “unknown”.
• Mask personal data you encounter as [REDACTED:{type}].
• Reject unsafe or unsupported tasks with a clear reason.
• For high risk outputs, set approval_required=true in JSON.

Role and audience switches

Add one of these to the User section to tune tone and constraints.

• Marketing: plain English, no hype, clear CTA, sources for stats.
• Sales: concise, next steps, no pricing promises without quote.
• Support: empathetic, step-by-step, no legal or medical advice.
• Executive: 5 bullets max, risks and decisions first, numbers with sources.

12 universal templates

Rewrite – 1) Tone-safe rewrite for external audience

System: You are a careful editor. Do not invent facts. Mask PII as [REDACTED:{type}].
User: Rewrite for an external audience. Tone = clear and respectful. No new claims. If numbers appear, add Sources at the end.
Input:
[PASTE TEXT]
Output JSON using the Rewrite schema: {"text","notes"}.

Rewrite – 2) SEO rewrite with entity emphasis

System: Emphasize entities and relationships, not keyword stuffing.
User: Rewrite to clarify the main entity and 3 to 6 attributes. Keep headings scannable and add 1 comparison table suggestion in "notes".
Input:
[PASTE TEXT]
Output JSON: {"text","notes"}.

Rewrite – 3) Policy-compliant rewrite

System: Remove prohibited phrases and risky claims.
User: Rewrite the text to comply with brand and legal policy. If policy conflicts, explain in "notes".
Input:
[PASTE TEXT]
Prohibited phrases: ["guaranteed", "100% risk-free"]
Output JSON: {"text","notes"}.

Summarize – 4) Executive 5-bullet summary

System: Summarize conservatively. No new claims.
User: Create an executive summary in ≤5 bullets. If any number appears, add Sources. Include top 1 to 2 risks.
Input:
[PASTE TEXT]
Output JSON Summary schema: {"bullets","actions":[],"risks"}.

Summarize – 5) Action items from meeting notes

System: Extract tasks only if explicitly mentioned.
User: From the notes, output concise action items assigned to an owner due within 14 days. If owner missing, set "owner":"unspecified".
Input:
[MEETING NOTES]
Output JSON Summary schema: {"bullets":[],"actions":[...],"risks":[]}.

Summarize – 6) Customer email response plan

System: Be helpful and safe. No promises or compliance claims.
User: Summarize the customer email and propose 2 response options with next steps. Do not write the email.
Input:
[CUSTOMER EMAIL]
Output JSON Summary schema: bullets = recap, actions = next steps, risks = potential pitfalls.

Extract – 7) Entities and attributes

System: Extract entities precisely. If unsure, set value="unknown".
User: Identify tools, people, products, metrics, and dates. Include attributes where present.
Input:
[PASTE TEXT]
Output JSON Extraction schema: {"entities":[{"type","value"}],"fields":{}, "confidence"}.

Extract – 8) Business fields with validation

System: Validate formats with regex where possible.
User: Extract customer, product, and amount (number). If amount not explicit, set to 0 and add note.
Input:
[PASTE TEXT]
Output JSON: {"entities":[],"fields":{"customer","product","amount"},"confidence"}.

Extract – 9) Citations for numeric claims

System: Require a source label for any number.
User: Extract all numeric claims with a short citation label (e.g., "vendor doc", "press release"). If no source, set citation="unknown".
Input:
[PASTE TEXT]
Output JSON:
{"entities":[],"fields":{},"confidence":0.0,"claims":[{"text":"string","citation":"string"}]}.

Classify – 10) Support ticket routing

System: Output JSON ONLY using the Classification schema.
User: Classify into one of ["billing","technical","sales","spam"]. If confidence < 0.8 return label="review".
Input:
[TICKET TEXT]
Output: {"label","confidence","reason"}.

Classify – 11) Priority and sentiment

System: Use explicit signals to avoid guessing.
User: Classify priority into ["p1","p2","p3"] and sentiment into ["neg","neu","pos"]. If missing signals, choose "p3" and "neu".
Input:
[EMAIL OR NOTE]
Output: {"label":"p1|p2|p3","confidence":0.0,"reason":"why","sentiment":"neg|neu|pos"}.

Classify – 12) Risk and approval routing

System: Set approval_required=true for high risk classes.
User: Classify content risk into ["low","medium","high"] based on legal claims, pricing promises, or PII presence.
Input:
[CONTENT]
Output:
{"label":"low|medium|high","confidence":0.0,"reason":"string","approval_required":true|false}.

Validation helpers and pass-fail checks

• JSON strictness: reject outputs with unknown keys or missing required keys.
• Numbers need sources: if any digit appears, require a “Sources” line or a citations array.
• Length limits: cap input length to reduce latency and costs.
• Approval flags: route high risk outputs for human review before publishing or syncing.

Rollout plan for teams

1. Pick 1 pattern per team and ship it as a template this week.
2. Add logging: model version, latency, cache hit, and pass or fail checks.
3. Run a 20-case weekly eval to catch drift and improve prompts.
4. Automate only after editors are comfortable with pass rates and guardrails.

FAQ

Do these templates work with any model?

Yes. They are model-agnostic. Keep schemas and guardrails identical so you can compare models fairly.

How do we keep outputs consistent across teams?

Store prompts and schemas in versioned files, reuse the same validation, and run the same weekly eval set across teams.

What if a template starts failing?

Inspect the top 5 failures, tighten rules, or add examples. Roll back to the last passing version until fixes are validated.

Final thoughts

Reusable prompt patterns accelerate adoption while reducing risk. Keep schemas strict, guardrails simple, and evaluations regular. Your content stays consistent, your automations remain stable, and your teams move faster with confidence.

Further reading on AIToolsBusiness:

• Evaluations & Guardrails – reduce hallucinations with tests and filters.