Prompt Design Workshop — Recipe & Exercises

Welcome & Purpose

This workshop helps teams rapidly learn prompt design patterns that produce repeatable, trustworthy outputs for real work. It focuses on practical recipes—persona prompts, context shaping, few-shot patterns, prompt chaining, and evaluation tests—so teams leave with tested prompts and a clear plan to pilot assistants in their workflows.

Intended Audience

Cross-functional teams, knowledge workers, product owners, support leads, engineers, and anyone who will prompt or supervise AI assistants.

Outcomes (what participants will be able to do)

• Explain the components of a high-value prompt and common failure modes.
• Design persona and system prompts that align with role and audience.
• Create context-shaping patterns and few-shot examples for consistent outputs.
• Construct simple chained workflows (call sequences) and basic tests for reliability.
• Apply evaluation metrics and a deployment checklist to pilot assistants safely.

Workshop Logistics (Recommended Formats)

Adaptable to a 90-minute sprint or a 3-hour deep workshop. Below is a recommended 120-minute agenda for a cross-functional team of 6–12 people.

1. 15 min — Intro, objectives, and safety basics (data, privacy, hallucinations)
2. 20 min — Primer: prompt anatomy & common patterns (persona, context, examples, constraints)
3. 35 min — Exercise A: Design a role persona & instruction set (paired work)
4. 20 min — Exercise B: Context shaping + few-shot examples (team test cases)
5. 15 min — Evaluation: run tests, score outputs, and iterate
6. 10 min — Next-step playbook: deployment checklist and responsibilities
7. 5 min — Close & commitments

Materials & Setup

• Projector or shared screen and access to a model playground (or offline examples)
• Templates (persona, context, few-shot) printed or in shared doc
• Example data or documents participants care about (specs, tickets, emails)
• Timer, whiteboard or collaborative document, and a simple scoring sheet

Facilitator Script & Guidance

Begin by asking: "What would a helpful assistant do for your role?" Capture 2–3 concrete jobs-to-be-done. Emphasize that prompts shape behavior and expectations, and that testing early reduces brittle assistants.

Prompt Anatomy (brief primer)

• System / Instruction — high-level role, constraints, and non-negotiables
• Context — relevant data, examples, available tools, or document excerpts
• Task / User Input — the specific user request or question
• Few-shot Examples — input-output pairs that demonstrate desired format and reasoning
• Output Constraints — format, length, style, and safety limits

Exercise A — Persona Design (paired, 20–30 min)

Goal: Create a concise persona + instruction block that produces behavior aligned to a team role.

1. Pick a target role (e.g., Customer Support Specialist, QA Engineer, Product Analyst).
2. Draft a 2–4 sentence system prompt defining persona, responsibilities, tone, and hard constraints (e.g., "Do not invent facts; cite sources when available").
3. Write one example user request and the ideal assistant response (these become few-shot examples).
4. Swap with another pair and test by running the prompt once; note mismatches.

Sample Persona Prompt

System: You are a Customer Support Assistant who prioritizes empathy and accurate troubleshooting. Always ask one clarifying question if the user request is ambiguous. Do not guess passwords or personal data. Provide step-by-step solutions and a short summary of next steps.

Exercise B — Context Shaping & Few-Shot (team, 20–30 min)

Goal: Provide the assistant with context and 2–3 example Q&A pairs so outputs match format and tone.

1. Collect 2–3 real inputs the team needs automated (ticket examples, research questions, routine reports).
2. For each, craft an example input and the desired output. Keep examples concise and representative.
3. Assemble: System prompt + context snippet + few-shot examples + task. Run and compare to desired output.
4. Iterate: adjust constraints, add one clarifying example, or reduce irrelevant context.

Few-Shot Pattern Example

Example 1:
  Input: User reports app crash when submitting form X.
  Output: Troubleshooting steps (3 bullets), probable cause, logs to collect.

  Example 2:
  Input: User asks how to reset API tokens.
  Output: Short steps with security note and link to internal KB.
  

Evaluation Metrics & Test Cases

Use these lightweight, practical metrics to decide whether a prompt is fit for pilot:

• Accuracy / Faithfulness — factual correctness, absence of hallucination (scale 1–5).
• Usefulness / Actionability — does the output help the user complete work (1–5).
• Format Compliance — respects required format, templates, and length (Yes/No).
• Robustness — performs acceptably across 3 variant phrasings (Pass/Fail).
• Safety & Data Handling — avoids exposing sensitive info or making unsafe recommendations (Pass/Fail).

Run each prompt against 5 representative inputs and record the scores. Aim for median ≥4 on usefulness and accuracy for pilot readiness.

Quick Troubleshooting Checklist

• Outputs are inconsistent — add one more few-shot example and tighten constraints.
• Assistant hallucinates facts — require citations, add "Do not invent" constraint, or provide reliable context sources.
• Responses too long or verbose — add explicit length and bullet-point constraints.
• Fails on edge cases — create targeted test cases and add guardrails (e.g., "If unknown, ask to verify").

Next-Step Playbook for Deployment

1. Choose a small pilot team and 1–3 concrete tasks where the assistant will help.
2. Document prompt versions, test cases, and scoring results.
3. Define monitoring: who reviews outputs daily and how errors are reported.
4. Plan a 2-week pilot: collect 50–100 interactions, measure metrics, and iterate weekly.
5. Enforce safety rules and data governance — restrict PII, log inputs, and maintain an audit trail.
6. Decide success criteria and scale plan (roles, integration points, KPIs).

Facilitator Tips & Common Pitfalls

• Keep examples focused — too many heterogeneous examples confuse the model.
• Prefer quality over quantity for context: include only what affects the output.
• Be explicit about disallowed behavior (e.g., inventing dates, giving medical advice).
• Encourage teams to treat prompts as living artifacts to be versioned and tested.

Closing & Reflection Prompts

Ask participants to write a one-line commitment: the prompt they will pilot and the metric they will track. Capture these in a shared doc and schedule a 1-week review.

Appendix — Example Prompts & Patterns

System + Task Example

System: You are a concise product analyst. Always show assumptions and cite the data source.
  Context: [insert short data snapshot or link to dataset]
  Task: Given this data, provide 3 possible hypotheses for the trend, with an actionable next test (one sentence each) and confidence level (low/medium/high).

Chained Call Pattern

1. Call 1: Extract structured facts from raw text (key fields).
2. Call 2: Use extracted facts + persona prompt to generate deliverable (email, summary, report).

Safety & Governance Notes

Include mandatory guardrails for any pilot: do not return PII unless explicitly redacted, log all outputs for review, and require human review for decisions with material risk. Train participants on when to escalate.