Team Prompting: Workshop, Maintenance and Benchmarking

Run a prompting workshop with your team

To become a reference, a prompting course must also help a team improve together. The goal is not for everyone to leave with a few isolated tips, but for the group to create a shared language, quality criteria and prompts that work in real workflows.

A strong prompting workshop starts from a concrete business case: customer support, document analysis, marketing production, meeting summaries, recruiting, code or reporting. The facilitator’s role is to frame the problem, make the group test several prompt versions, then turn the best results into reusable templates.

Recommended format for a 90-minute session

1. Choose a shared use case and define the expected result.
2. Show an intentionally vague prompt, then identify its weaknesses.
3. Build an improved version with goal, context, constraints, format and criteria.
4. Have participants work in pairs on two variants of the same prompt.
5. Compare the answers with a simple rubric: accuracy, usefulness, structure, risks and reusability.
6. Document the best version in the team prompt dossier.
7. Decide who will test the prompt in real conditions and when it will be reviewed.

Deliverables to produce during the workshop

• One reference prompt validated by the group.
• Two variants adapted to different complexity levels.
• A short rubric for evaluating AI answers.
• A list of data that should be excluded from the team’s prompts.
• An owner responsible for maintaining the prompt after the first tests.

Tip: start with a frequent but low-risk use case. A team learns better when it can quickly compare several outputs without exposing sensitive data or depending on a critical decision.

---

Update your prompts as models evolve

AI models change quickly. A prompt that works today may become too long, too constrained or simply less adapted after a model update. The durable skill is not memorizing a perfect prompt, but knowing how to revise it.

When to review a prompt

• when the model you use changes;
• when the result becomes less stable;
• when your business context or audience evolves;
• when the prompt contains outdated examples, tools or constraints;
• when you want to share it with a team.

Revision checklist

1. Review the goal: does the prompt still help produce the right decision or deliverable?
2. Remove unnecessary constraints: some may have been useful with an older model but no longer are.
3. Test the same prompt on a recent case and compare the result with your success criteria.
4. Add a limit or verification rule if the topic is sensitive.
5. Record the prompt version, model used and last test date.

This discipline prevents your prompt library from becoming an unusable archive. It helps you keep only prompts that still produce reliable results in your real context.

Tip: schedule a short review of your important prompts every two or three months, or after a major model or workflow change.

---

Frequently Asked Questions

---

Prompt benchmark: compare before adopting

A reference-level course should not encourage people to choose a prompt because it “looks better”. For important use cases, you need to compare several versions on the same test cases, with explicit criteria and a version decision.

A prompt benchmark can stay simple. Prepare five to ten representative examples, test each prompt version on those examples, then score the answers with a stable rubric. The goal is not to produce a perfect measure, but to make improvement visible and debatable.

Lightweight benchmark protocol

1. Choose five to ten real cases covering frequent, difficult and edge situations.
2. Define what a good answer must include and what it must avoid.
3. Test the current prompt version, then one or two variants.
4. Score each output with the same criteria: accuracy, usefulness, structure, constraint following and risks.
5. Identify cases where a variant fails even if it looks better on one isolated example.
6. Keep the winning version with its date, score and remaining fragile cases.

Simple criteria to use

• Accuracy: does the answer respect the facts and provided context?
• Usefulness: can the answer be used without heavy rewriting?
• Structure: is the requested format followed?
• Robustness: does the prompt work on ambiguous or incomplete cases?
• Safety: does the answer avoid risky actions, data or claims?

Tip: a prompt that wins on one example can fail on an edge case. Always keep a few difficult cases in your benchmark.

---

Going further

These skills are built through practice. Resume the structured path in the free prompting course, and train with the interactive exercises: optimization and iteration, multi-persona debate.