Prompt Engineering Workshop
3-Hour Prompt Engineering Workshop:
Course Overview
This 3-hour interactive workshop equips participants with core prompt engineering skills. Through discussions, hands-on exercises, and real-world scenarios, you’ll learn to craft effective prompts and get better results from AI models.
Duration: 180 minutes
Audience: Practitioners using Claude, GPT, Gemini, Grok, and other frontier models
Core Shift: From “clever wording” to outcome specification + context engineering. Reasoning is now largely a dial you set, not a sentence you write.
Workshop Overview
1. Workshop Kickoff & Mindset Shift (10 min)
- Quick participant check-in: current AI use cases and biggest frustrations.
- Core message: In 2026 the highest-leverage skill is no longer “writing clever prompts.” It is specifying the desired outcome clearly and curating the right context.
- Live demo of the same vague request producing widely different quality depending on structure and context.
- Workshop agreement: Treat every first output as a draft. Iteration is the real skill.
2. Prompt Engineering in 2026: From Prompts to Context (12 min)
- Evolution: Classic prompt engineering (wording) → Context engineering (what the model sees: system instructions, retrieved knowledge, conversation history, tool results, memory).
- Why the shift happened: Reasoning models + long context + agents made pure instruction-tweaking less dominant.
- What still matters most: Clear goals, success criteria, relevant context, examples for format, and structured output contracts.
- Practical definition for this workshop: Designing the full information environment so the model can deliver reliable, high-value results.
3. Building a Fast Prompt Practice (10 min)
- Systems that compound speed:
- Personal prompt library / snippet system
- Reusable templates (Role + Goal + Context + Format + Constraints)
- Voice dictation for thinking out loud
- Model-specific projects / custom instructions / system prompts
- “Draft → Diagnose → Refine” as the default workflow instead of perfection on the first try.
- Quick tip: Version prompts like code when they matter in production.
4. How Modern LLMs Actually Process Input (12 min)
- Practical mental model (not deep architecture):
- Next-token prediction under uncertainty
- Context window as limited working memory
- Attention dilution: important instructions can get lost in noise
- Why models still hallucinate and how good context reduces it
- Reasoning models vs standard models: internal thinking budget vs external step-by-step instructions.
- Temperature / sampling still matter, but the biggest new lever is the thinking/effort dial.
5. Outcome-First Prompting (Steering vs Over-Commanding) (12 min)
- 2026 best practice: Describe the destination and success criteria more than the exact procedure.
- Reasoning models perform better when given high-level goals and constraints; excessive step-by-step instructions can hurt.
- Classic models still benefit from clearer procedural guidance.
- Examples of over-commanding that make outputs rigid or lower quality versus clean outcome specification.
- Rule of thumb: Match the level of detail to the model type.
6. Context Crafting: Role, Scene, Audience & Success Criteria (15 min)
- Highest-ROI elements to include:
- Precise role or perspective (only when it adds real signal)
- Audience and purpose
- Relevant background / constraints the model cannot know
- Explicit success criteria (“Done means…”)
- Preferred tone, length, and format
- Structure techniques that work across models: clear sections, XML-style tags (especially strong with Claude), Markdown headers.
- Exercise: Rewrite a vague request into a high-signal context block and compare results.
7. Teaching by Example: Few-Shot Prompting (15 min)
- Still one of the strongest techniques for controlling format, style, and edge-case handling.
- Quality > quantity: 2–5 carefully chosen, consistent examples beat long lists.
- Best used for shape (output structure, tone, classification boundaries) rather than teaching complex reasoning on modern models.
- Include at least one near-boundary or edge-case example when reliability matters.
- Live exercise: Build a short few-shot prompt for a real participant task.
8. Reasoning Control & Thinking Effort (15 min)
- Major 2026 change: Reasoning depth is now largely a control (effort / thinking level / budget), not a phrase you append.
- When to use explicit Chain-of-Thought still: non-reasoning models or when you need visible intermediate steps for verification.
- When to avoid forced “think step by step”: most frontier reasoning models — it can add latency and noise.
- Practical guidance: Set the effort dial to match task difficulty, then focus the prompt on outcome and constraints.
- Brief mention of related patterns (self-consistency, multi-path exploration) and when they are worth the cost.
9. Structured & Reliable Outputs (12 min)
- Why structure is non-negotiable for anything that feeds tools, pipelines, or further automation.
- Native structured output / JSON schema support (preferred) versus prompt-enforced formats.
- Techniques: schema definition, XML tags, Markdown tables, prefilling the start of the response, validation + retry loops.
- Always pair structure with a clear “respond only in this format” instruction when using pure prompting.
- Demo: Turning free-form answers into clean, parseable outputs.
10. Constraints That Actually Work (10 min)
- Positive constraints outperform heavy negative lists in most cases.
- Effective patterns: length limits, required elements, style boundaries, scope definitions.
- Negatives are still useful for hard prohibitions (“never invent citations,” “do not include marketing language”).
- Risk of over-constraining: models become brittle or refuse useful creativity.
- Framing tip: Convert “don’t do X” into “do Y instead” whenever possible.
11. The Iteration & Evaluation Loop (15 min)
- The real skill: systematic diagnosis.
- Diagnosis questions: Missing context? Ambiguous success criteria? Wrong examples? Conflicting instructions? Model type mismatch?
- Techniques: critique-and-revise, progressive disclosure of requirements, “what’s still missing?”, side-by-side comparison.
- Lightweight evaluation: keep a small set of real test cases and re-run after changes.
- Live group iteration: Start with a mediocre prompt and improve it in 3–4 rounds.
12. Collaborative Discovery Prompting (Interview Style) (10 min)
- Pattern: Ask the model to interview you first until it has enough information.
- Highly effective for ambiguous or complex projects (strategy, specs, learning plans, creative briefs).
- Control levers: limit number of questions, request prioritization, allow the model to propose assumptions.
- Demo of turning a vague idea into a high-quality brief through guided questions.
13. Advanced Levers & Model Differences (12 min)
- System / developer messages vs user messages — persistent rules belong in the system layer.
- Key parameters: temperature, top_p, and especially reasoning/thinking effort controls.
- Model family differences (high-level):
- Claude: Strong with XML structure and extended thinking
- OpenAI reasoning models: Prefer clear goals over detailed procedures
- Others: Vary in long-context handling and formatting preferences
- Emerging patterns: prompt chaining, simple agent loops, self-critique, combining few-shot + structured output + effort control.
- When pure prompting is no longer enough (tools, retrieval, memory, multi-agent).
14. Common Pitfalls & Closing (20 min)
Frequent 2026 anti-patterns:
- Treating reasoning models like 2023 models (forcing step-by-step when the dial exists)
- Overloading a single prompt with everything
- Ignoring context quality and only polishing wording
- No success criteria or evaluation cases
- Assuming long conversations retain perfect memory of early instructions
- Copy-pasting prompts across models without adaptation
- Heavy negative constraint lists instead of positive framing
- Skipping iteration after the first decent result
15. Closing
- Key takeaways recap
- Recommended practice: Keep a prompt journal + small personal eval set for 2 weeks
- Resources for going deeper (context engineering, evaluation tools, model-specific guides)
- Open Q&A