You asked for the real, engineered stuff — prompts written the way ChatGPT-class models want them: explicit role, clear objective, tight constraints, deterministic output formats, and built-in verification. Below are 10 production-grade templates you can drop into your workflow. Each includes:

• Message structure (System + User)
• Variables you replace in [brackets]
• Output schema so responses are parseable
• Guardrails to reduce hallucination
• Self-check so the model validates its own work.

0) Base Prompt Spec (use for all)

Pattern

• Role: who the model is
• Objective: single verb + artifact
• Inputs: precise, labeled blobs
• Constraints: length, style, locale
• Output format: strict JSON schema or Markdown table
• Guardrails: what not to do, when to ask questions
• Self-check: fast validation pass, repair if needed

1) Meeting → Decisions, Actions, Risks (DAR)

System

You are a senior project coordinator. Be precise, concise, and action-oriented.

User

OBJECTIVE: Extract decisions, actions, and risks from a meeting transcript.

INPUT_TRANSCRIPT: """[paste transcript]"""

CONSTRAINTS:
- Use present tense; max 14 words per bullet.
- Owners must be real names from transcript, else "TBD".
- Deadlines in ISO date, else "TBD".

OUTPUT_FORMAT: JSON exactly matching this schema:
{
  "decisions": [{"summary": "..."}],
  "actions": [{"owner": "...", "task": "...", "deadline": "YYYY-MM-DD"}],
  "risks": [{"risk": "...", "mitigation": "..."}],
  "followups": [{"question": "...", "owner": "..." }],
  "rationale": "≤40 words on how items were derived"
}

GUARDRAILS:
- If the transcript lacks owners or dates, use "TBD".
- Do not infer facts not mentioned.

SELF_CHECK:
1) Ensure JSON is valid and fields exist.
2) No action bullet exceeds 14 words.
If failed, repair and re-emit JSON.

Why it works: fixed schema + word caps → clean, scannable minutes with zero post-edit.

2) Email Reply Composer with Tone & CTA

System

You are an executive communications assistant. Write crisp, courteous business emails.

User

OBJECTIVE: Draft a reply email.

INPUT_EMAIL: """[paste inbound email]"""
CONTEXT: recipient_role="[role]", relationship="[cold|warm|internal]"

CONSTRAINTS:
- 110–140 words; 1 CTA; no more than 3 sentences per paragraph.
- Tone: [friendly|formal|assertive], Region: [US/UK/IN] English.

OUTPUT_FORMAT: JSON
{
  "subject": "...",
  "body": "single string with line breaks",
  "cta": "...",
  "rationale": "≤30 words"
}

GUARDRAILS:
- Do not promise timelines or pricing unless provided.
- If key info is missing, include 1 clarifying question before CTA.

SELF_CHECK:
- Body length within 110–140 words.
- Exactly one CTA present.

3) Source-Bound Research Brief (Anti-Hallucination)

System

You are a research analyst. Only use sources provided by the user.
If uncertain, say "Unknown based on provided sources".

User

OBJECTIVE: Produce a 1-page research brief.

SOURCE_PACK:
1) [title]: [url or pasted excerpt]
2) [title]: [url or pasted excerpt]
3) ...

CONSTRAINTS:
- 120–150 words executive summary, then 3 sections: Key Stats, Insights, Open Questions.
- Every stat must have a source_id.

OUTPUT_FORMAT: JSON
{
  "summary": "...",
  "key_stats": [{"stat": "...", "source_id": 1}],
  "insights": [{"insight": "...", "source_id": [1,2]}],
  "open_questions": ["..."],
  "citations": [{"source_id": 1, "title": "...", "url": "..."}]
}

GUARDRAILS:
- No external knowledge; no unsourced numbers.
- If a figure appears in multiple sources and conflicts, pick none; add to open_questions.

SELF_CHECK:
- Each stat has a valid source_id.
- No numbers appear in summary without a source in key_stats.

4) Data → Insights → Recommendations (Table-Safe)

System

You are a data analyst. Prefer simple language and concrete recommendations.

User

OBJECTIVE: Turn dataset into insights.

DATASET: CSV columns: [list]  ROWS: """[paste table or sample]"""
BUSINESS_GOAL: "[e.g., reduce CAC by 15%]"

CONSTRAINTS:
- No math beyond descriptive stats unless explicitly present.
- 3 insights, 2 trends, 3 recommendations tied to BUSINESS_GOAL.

OUTPUT_FORMAT: JSON
{
  "insights": [{"text":"...", "evidence":"column=value", "impact_estimate":"low|med|high"}],
  "trends": [{"metric":"...", "direction":"up|down", "period":"...", "evidence":"..."}],
  "recommendations": [{"action":"...", "why":"...", "expected_effect":"..."}],
  "assumptions": ["..."]
}

GUARDRAILS:
- If data insufficient, state assumption plainly in "assumptions".
- No causal claims.

SELF_CHECK:
- Each recommendation references at least one insight or trend.

5) Campaign Ideation with Impact Scoring

System

You are a growth strategist. Optimize for impact under constraints.

User

OBJECTIVE: Generate campaign ideas.

CONTEXT:
- Product: "[desc]"
- Audience: "[ICP]"
- Channel constraints: "[e.g., organic only]"
- Budget: "[number or 'zero']"
- Timeframe: "[e.g., 4 weeks]"

SCORING_WEIGHTS: impact=0.5, cost=0.2, speed=0.2, risk=0.1

OUTPUT_FORMAT: JSON
{
  "ideas": [
    {
      "title":"...",
      "concept":"...",
      "steps":["..."],
      "score": {"impact":0-10,"cost":0-10,"speed":0-10,"risk":0-10,"weighted_total":0-100}
    }
  ],
  "top_pick_reason": "≤40 words"
}

GUARDRAILS:
- Respect constraints strictly.
- If a metric is unknown, assign neutral 5 and explain.

SELF_CHECK:
- weighted_total = sum(metric*weight)*10 (round to integer).
- Steps ≤ 5 per idea.

6) Slide Deck Blueprint (Slide-by-Slide)

System

You are a presentation architect. Create concise slide blueprints.

User

OBJECTIVE: Convert content into a 6-slide outline.

INPUT_CONTENT: """[paste doc or notes]"""
AUDIENCE: "[execs|customers|technical]"
GOAL: "[approve budget|educate|sell]"

CONSTRAINTS:
- 6 slides exactly: Title, Problem, Evidence, Solution, Impact, Next Steps.
- ≤4 bullets per slide; ≤12 words per bullet.

OUTPUT_FORMAT: JSON
{
  "slides": [
    {"title":"...", "bullets":["..."], "visual_hint":"icon|chart|screenshot"}
  ],
  "notes":"≤50 words on narrative flow"
}

GUARDRAILS:
- No new claims; only reorganize provided content.
- If evidence missing, add a placeholder bullet "Evidence needed: [specify]".

SELF_CHECK:
- Exactly 6 slides; bullet limits satisfied.

7) Policy Simplifier with Compliance Flags

System

You are a compliance communicator. Simplify without losing obligations.

User

OBJECTIVE: Rewrite policy in plain language.

POLICY_TEXT: """[paste]"""
REGION: "[US|EU|IN]"  READING_LEVEL: "Grade 8"

CONSTRAINTS:
- 200–250 words; sections: Purpose, Rules, Deadlines, Penalties, Contacts.
- Bold deadlines and penalties with ** **.

OUTPUT_FORMAT: Markdown with the five sections as H3 headings.

GUARDRAILS:
- Do not add or remove obligations; if ambiguous, mark "(clarification required)".
- Preserve numeric thresholds exactly.

SELF_CHECK:
- All numbers/dates match source.
- Each section present.

8) Task Prioritization using Eisenhower + Dependencies

System

You are a productivity coach. Classify and sequence tasks.

User

OBJECTIVE: Prioritize tasks and propose a schedule.

TASK_LIST:
- [task] | owner=[name] | effort=[S|M|L] | due=[YYYY-MM-DD or TBD] | dep=[task id or none]

CONSTRAINTS:
- Classify using Eisenhower (UI/NI vs. IN/NN).
- Propose a 5-day schedule (8h/day) honoring deps and due dates.

OUTPUT_FORMAT: JSON
{
  "matrix": {"urgent_important":[...],"urgent_not_important":[...],"not_urgent_important":[...],"not_urgent_not_important":[...]},
  "schedule":[{"day":"YYYY-MM-DD","blocks":[{"task":"...","owner":"...","duration_h":...}]}],
  "tradeoffs":"≤50 words"
}

GUARDRAILS:
- If overbooked, push NI first and explain in tradeoffs.

SELF_CHECK:
- No day exceeds 8h; dependencies respected.

9) FAQ Builder with Beginner/Advanced Split

System

You are a documentation writer. Separate content by audience level.

User

OBJECTIVE: Create FAQs from product brief.

PRODUCT_INFO: """[paste key info]"""

CONSTRAINTS:
- 5 beginner FAQs; 5 advanced FAQs.
- Answers ≤60 words; plain English; link anchor placeholders like [Learn more].

OUTPUT_FORMAT: JSON
{
  "beginner":[{"q":"...","a":"..."}],
  "advanced":[{"q":"...","a":"..."}]
}

GUARDRAILS:
- If info missing, add question to "advanced" with an honest limitation note.

SELF_CHECK:
- Count totals; word limits.

10) Critic-Then-Rewrite Guardrail (Drop-in Post-processor)

Use this after any generation to raise quality.

System

You are a meticulous editor. You critique first, then improve.

User

OBJECTIVE: Critique and rewrite the prior output to meet strict standards.

INPUT: """[paste model's previous output]"""

QUALITY_BAR:
- Accuracy over flourish; eliminate vague claims.
- Short sentences; remove filler.
- Keep structure and schema intact.

OUTPUT_FORMAT: JSON
{
  "critique": ["bullet issue 1","bullet issue 2","..."],
  "revised": { /* same schema as INPUT, corrected */ }
}

GUARDRAILS:
- Do not invent facts; if unknown, mark "Unknown".
- Preserve required fields and formats.

SELF_CHECK:
- Revised output validates against schema; all issues addressed.

Implementation Tips That Make These Work Reliably

1. Pin the schema: Models behave best with explicit keys and short, stable schemas.
2. Cap bullets/words: Length caps reduce waffle and make outputs scannable.
3. Force uncertainty handling: Require "Unknown" or "TBD" instead of fabrication.
4. Require citations or source_ids when summarizing external info.
5. Add a self-check step: Ask the model to validate JSON and constraints before returning.
6. Version your prompts: Keep v1.3 etc. so teams know which template they use.
7. Few-shot only where precision matters: Insert 1–2 compact, correct exemplars if you see drift.
8. Separate creation from verification: First generate, then run #10 critic pass.

Copy-Paste QA Snippet (use with any JSON output)

SELF_CHECK:
- Attempt to parse the output as JSON. If parsing fails, fix formatting and re-emit.
- Ensure all required keys exist and no extra keys are included.
- Validate enumerations and date formats.

Use these templates as building blocks. Swap in your variables, keep the schemas tight, and let the guardrails do the heavy lifting. This is the difference between “nice demo” and daily, dependable time savings.