Measuring Prompt Engineering Skills: Criteria, Exercises and Mistakes to Avoid

“Write a good prompt.” The instruction looks simple. Yet it raises a major problem for anyone who has to assess a training course: what is a good prompt? A long text? A structure with a role, a context and a format? A convincing answer on the first try? None of these signs is enough to demonstrate a professional skill.

An employee can write a very detailed prompt and produce an unusable result. Another can write a short instruction, spot an error, correct their method and deliver a reliable document. Which of them really masters prompt engineering?

Prompt engineering is less about producing a perfect formula than about steering an interaction with an AI system to accomplish a task. The employee must understand the need, choose the useful information, specify the expected result, control the data sent, test the answer, detect flaws and decide whether the deliverable can be used. The prompt is only a trace of this process.

This distinction changes the whole assessment. If you mark only the form of the instruction, you easily reward verbosity or mechanical use of a template. If you observe the complete work, you can measure a skill that is useful to the company: obtaining a reliable, explainable and reusable result under controlled conditions.

Why the Apparent Quality of a Prompt Is Not Enough

A well-written prompt can produce a bad result

An instruction can contain a role, six constraints and an output format without meeting the real need. For example, asking an AI to write a highly structured sales analysis guarantees neither the quality of the data, nor the accuracy of calculations, nor the relevance of recommendations. Structure sometimes improves the answer, but it does not replace business judgement or verification.

The documentation of the main vendors converges on one point: you must define success criteria and test outputs. Anthropic places defining measurable criteria and building evaluations at the heart of the prompt engineering cycle. Google, for its part, reminds us that prompt design is an iterative process based on observed results rather than an unchanging recipe.

An elegant answer can hide an error

Generative models produce fluent text, including when the information is incomplete or false. An assessment that rewards writing quality without requiring a fact check mostly measures the model’s ability to convince. In an HR, legal, financial, sales or technical task, this bias can turn an impressive demonstration into an operational risk.

The skill depends on the business context

The same employee may be autonomous at rephrasing an internal email and need strict supervision to analyse a contract or produce a financial recommendation. The expected level therefore depends on the task, the audience, the sensitivity of the data, the impact of an error and the available means of verification.

The European Commission follows this contextual logic for AI literacy. Its interpretation of Article 4 of the AI Act asks organisations to take into account people’s knowledge, experience and training, the context of use and the risks of the system. It also specifies that the text does not require measuring a given level individually or obtaining a certificate. Assessing nonetheless remains useful to steer skills development and show that the actions chosen respond to real uses. For the full regulatory framework, see our article on what the AI Act really requires of companies.

What Should You Actually Measure?

A robust grid must cover the whole work chain. Eight dimensions make it possible to observe prompt engineering without reducing it to writing an instruction.

Dimension Assessment question Observable evidence
— — —
1. Framing the need Does the employee understand the task, the recipient and the expected decision? Restatement of the need, questions asked, expected result made explicit.
2. Selecting context Do they provide the useful information without overloading or exposing unnecessary data? Selected corpus, data removed or anonymised, assumptions flagged.
3. Building the instruction Does the request specify the action, constraints, format and criteria? Prompt that is understandable, suited to the task and usable by a third party.
4. Steering the interaction Can they break down, iterate and correct instead of starting over at random? Version history, justification of changes, stopping at the right time.
5. Verification Do they check facts, calculations, sources, omissions and inconsistencies? Checklist, reconciliation with sources, documented corrections.
6. Security and compliance Do they identify sensitive data and uses that require validation? Refusal of a risky share, anonymisation, choice of an authorised environment.
7. Quality of the deliverable Does the result meet the professional need and the stated criteria? Accurate, actionable deliverable suited to the audience and working format.
8. Transfer and reuse Does the method work on a similar but new task? Adaptation to a second case, documented template, limits explained.

This structure is consistent with DigComp 3.0, which defines digital competence as a combination of knowledge, skills and attitudes. The framework now integrates AI across all areas and insists in particular on the critical evaluation of information, sources and the processes that produce them. A full assessment therefore cannot be limited to memorising a prompting technique.

A 100-Point Prompt Engineering Assessment Grid

Weighting must reflect the risks and the value of the work. The following grid suits a common professional task, such as writing a proposal, analysing a corpus or creating a procedure. It can be adapted to the job.

Criterion Points Expected indicators
— — —
Framing the need 15 Objective, recipient, end use, constraints and missing questions correctly identified.
Choice of data and context 15 Relevant information, traceable corpus, no unnecessary or unauthorised data.
Quality of the instruction 15 Explicit task, suitable format, success criteria and limits clearly stated.
Iteration strategy 10 Relevant breakdown, targeted corrections, comparison of variants where needed.
Verification and critical thinking 20 Facts, calculations, sources, consistency and completeness checked using an explicit protocol.
Security, confidentiality and responsibility 10 Risks recognised, data protected, human intervention placed at the right point.
Quality of the professional deliverable 10 Usable, precise result suited to the target and meeting expectations.
Transfer and documentation 5 Reusable method, justified choices, limits and conditions of use documented.

Interpreting the score

  • 0 to 39 points — fragile practice: the person relies heavily on the model and does not secure the task enough.
  • 40 to 59 points — guided practice: they can complete simple tasks with an outline, a checklist and validation.
  • 60 to 79 points — autonomy on mastered uses: they frame, iterate and check low- or medium-criticality tasks correctly.
  • 80 to 100 points — confirmed mastery: they adapt their method, handle risks, justify their choices and transfer the approach to new cases.

Possible disqualifying rules. Whatever the score, the assessor may consider the skill not demonstrated if the participant knowingly sends confidential data to an unauthorised tool, uses an invented source without flagging it, hides a critical error or distributes a high-impact deliverable without the planned validation.

Building a Comparable and Fair Assessment

Give the same mission, the same corpus and the same constraints

Comparing two participants only makes sense if they have the same information. The test must specify the context, the audience, the deadline, the authorised data, the format of the deliverable and the success criteria. The assessor avoids marking knowledge of information a candidate did not have access to.

Keep traces of the process

The final deliverable does not show how the participant worked. Ask for a light evidence file:

  • the restatement of the need;
  • the prompt or prompts used;
  • one or two relevant intermediate versions;
  • the verification checklist;
  • the corrections made;
  • a short justification of decisions.

This trace distinguishes success obtained by chance from a reproducible method. It also avoids attributing all the quality of the result to the employee when it may depend on the model, the tool version or an example supplied.

Test several outputs when stability matters

A generative AI does not always produce the same answer. For recurrent use, a single try can give a misleading impression. The assessor can ask the participant to test the instruction on three variants of the same case: another customer, another format, a missing data point or an extra objection. The skill shows in the ability to anticipate gaps and maintain the quality level.

Adapt the weighting to the risk

For an internal post, writing quality can weigh more. For a regulatory summary, source traceability and caution must dominate. For data analysis, calculation and consistency checks are strengthened. Prompt engineering is never an abstract skill: it serves an activity whose consequences guide the assessment.

Four Practical Exercises to Assess the Skill

Exercise 1 — Turn a vague request into a usable brief

Situation: a manager says: “Write a convincing email to chase our customers.” The participant receives an incomplete company sheet and must prepare the mission before generating anything.

Expected work:

  1. identify the missing information;
  2. ask the requester five priority questions;
  3. formulate the success criteria;
  4. write a prompt suited to one of the possible answers;
  5. explain what must be checked before sending.

What the exercise measures: framing, refusing to fill gaps arbitrarily, adapting to the audience and the ability to define quality before generating. A competent participant does not rush into writing.

Exercise 2 — Produce a summary from a closed corpus

Situation: the participant receives three internal documents and must prepare a one-page decision note for a management committee. The AI must use only these sources.

Constraints: distinguish facts, hypotheses and missing information; tie every figure to its source; flag contradictions; propose two options without inventing data.

Evidence requested: initial prompt, citation rule, first output, verification table and corrected final note.

What the exercise measures: selecting context, grounding the answer, vigilance against hallucinations and quality of verification. NIST’s recommendations on generative AI give a central place to testing, evaluation, verification and validation processes. The company must therefore look at the control procedure, not only the delivered document.

Exercise 3 — React to a confidentiality risk

Situation: an instruction contains the names, contact details, pay and managerial comments on several employees. The participant must prepare an HR analysis using an assistant not approved by the company.

Expected response: stop the input, identify the data that must not be sent, propose an aggregated or anonymised version, check the authorised tool and keep human validation. The exercise can continue with a fictitious dataset.

What the exercise measures: judgement, security and the ability to change the process. A person who writes a technically excellent prompt but ignores the risk cannot be considered autonomous.

Exercise 4 — Transfer the method to a new task

Situation: after working on a sales email, the participant must use the same approach to build an internal procedure or analyse customer feedback.

Expected work: adapt the framing, change the criteria, select a different format, anticipate the risks specific to the new case and explain which elements of the first prompt must not be copied.

What the exercise measures: transfer, understanding of principles and autonomy. The person must not just reuse a “magic prompt”; they must rebuild an adapted method.

Worked Example: Assessing a Sales Reporting Prompt

Insufficient version

“Analyse these sales and give me a good report with recommendations.”

This request specifies neither the period, nor the decision expected, nor the reliable columns, nor the recipient, nor the calculations to perform. If the output looks correct, the assessor cannot tell whether it meets the need.

Structured version

“You are assisting the sales manager. From the attached file, compare third-quarter revenue with second-quarter revenue by range and by region. Flag missing data. Present the differences in value and percentage in a table, then propose three hypotheses to check. Do not present any hypothesis as a fact. End with four questions to ask regional managers.”

This version is more usable, but it still does not prove the skill. The participant must check the totals, the definition of periods, empty cells, duplicates and the consistency of percentages. They must also examine whether the recommendations really rest on the data. The final mark therefore depends as much on checking as on wording.

Debrief questions

  • Why did you choose these indicators?
  • What information was missing before the first try?
  • Which error in the output could have led to a bad decision?
  • What did you change between the two versions and why?
  • Under what conditions could this prompt be reused?

The Most Frequent Mistakes in Assessment

1. Marking the length of the prompt

A long prompt can contain contradictions, useless details or impossible constraints. A short instruction can be excellent if the context is already available. Mark useful precision, not the number of lines.

2. Imposing a single formula

“Role, context, task, constraints, format” frameworks are good teaching reference points. They must not become a mechanical marking scheme. Some tasks require examples, a corpus, tools, a sequence of sub-tasks or an evaluation protocol; others can be solved directly.

3. Confusing tool mastery with business skill

Knowing the menus of ChatGPT, Claude or Gemini does not guarantee that someone can analyse a customer need or interpret a financial table. The assessment must cross two dimensions: steering the AI and the quality of the professional decision.

4. Using a case that is too simple

A rephrasing instruction with no constraint or risk produces little evidence. A good case has an ambiguity, a missing piece of information, a format constraint, a point to verify and a decision to make.

5. Changing tool or model between participants

Results can vary depending on the model, its version, the available functions and the settings. For a summative assessment, standardise the environment as far as possible or document the differences precisely.

6. Keeping only the final result

Without a trace, it is impossible to know whether the participant detected an error or whether the model simply produced a good answer. The evidence file must stay light, but it must show the essential decisions.

7. Letting AI assign the mark alone

A model can help spot gaps in a grid, but it must not be the sole assessor. Its mark depends on the assessment prompt, the wording of criteria and its own biases. The human assessor keeps the decision, especially for security, business relevance and borderline cases.

8. Assessing without a second attempt

Skill also shows in the ability to learn from feedback. A good test provides short feedback, then a new attempt. Improvement between the two versions is precious evidence of understanding and transfer.

A Two-Hour Teaching Sequence Before the Final Assessment

Our AI programmes follow a “understand, produce, verify and reuse” logic. This progression can be applied to a short practice session.

Time Activity Output
— — —
0–15 min Individual diagnostic on a simple case, with no model supplied. Initial prompt and raw result.
15–35 min Group analysis of gaps: need, context, constraints, format, criteria. List of observable flaws.
35–55 min Commented demonstration of a complete approach. Framing outline and checklist.
55–85 min Guided practice on a corpus and a business deliverable. Two versions and a correction log.
85–105 min Peer review using the 100-point grid. Factual feedback from a peer.
105–120 min New attempt and commitment to transfer. Improved version and an on-the-job application task.

After the session, the manager can propose a real low-risk task within seven days. The employee submits the brief, the useful traces, the deliverable and their verification. This work confirms that the skill leaves the training room and settles into daily activity.

Measuring Progress Without Creating a Bureaucracy

The L&D dashboard can stay simple. Five indicators are often enough:

  • average score per dimension, to identify the need for remediation;
  • success rate on a new task, to measure transfer;
  • rate of critical errors detected before distribution;
  • time needed to reach a validated deliverable;
  • share of outputs accompanied by evidence of verification.

Time savings must not be isolated. A ten-minute reduction has no value if it increases errors, rework or the risk of disclosure. The right indicator combines efficiency, quality and risk control.

Prompt Template Ready to Adapt

Mission: from the file provided, produce [deliverable] for [audience] to enable [decision or action].

Constraints: use only authorised sources; flag missing information; protect sensitive data; respect [format, length, tone and deadline].

Work requested: restate the need, define your success criteria, choose the context elements, build and test your prompt, check the output, correct it, then justify your main decisions.

Evidence to submit: prompt(s), a useful intermediate version, verification checklist, final deliverable and a ten-line reflective note.

Assessment: 100-point grid, with particular attention to verification, security and transferability.

FAQ on Assessing Prompt Engineering

Should a perfect prompt be required on the first try?

No. In professional use, iteration is part of the job. Observe whether the person analyses the first output, changes the right parameters and stops when the criteria are met.

Can all jobs be assessed with the same grid?

The general structure can be common, but weightings and cases must change. A marketing team, an HR department and a financial controller use neither the same data, nor the same evidence, nor the same level of risk.

Is a multiple-choice test enough?

It can check certain knowledge: definitions, risks or good practice. It does not demonstrate the ability to frame a task, steer the AI, detect an error and deliver a professional result. A practical exercise is essential.

Should ChatGPT, Claude and Gemini be assessed separately?

For a transferable skill, it is better to assess the common principles first. Complementary modules can then check the functions specific to the working environment the company has chosen.

How can cheating be limited during the assessment?

Using AI is part of the test; the aim is therefore not to prevent it. Ask instead for the traces, an oral justification, an unexpected variant and a second task. Someone who simply copied a prompt will struggle to explain and transfer their method.

Does the AI Act require the company to run this test?

No. The Commission specifies that Article 4 requires neither a mandatory individual measurement of a level of knowledge nor a specific certificate. The company must nonetheless support AI literacy, taking into account its uses, its audiences and its risks. A practical assessment is a way of steering this effort, but it must remain proportionate.

How often should people be reassessed?

An initial assessment, evidence of transfer a few weeks after the training and a check when a tool or process changes significantly form a reasonable base. Teams exposed to sensitive uses may need more frequent exercises.

What is prompt engineering?

Prompt engineering consists of formulating and improving the requests sent to an LLM (large language model) to obtain a more precise final answer: providing examples (few-shot prompting), breaking down reasoning (chain of thought), specifying the output format. In a company, what is assessed is less the formula than the ability to steer artificial intelligence, check the answer and protect data.

What is the difference between a script and a prompt?

A script is a series of instructions executed deterministically; a prompt is a natural-language request given to an AI model, whose answer can vary from one attempt to the next. That is why we assess the method (framing, verification, iteration) more than the text of the prompt.

Assessing Your Teams’ AI Skills

Marmignon Brothers builds AI pathways from your tasks, your tools and your risks. We can adapt this grid to your jobs, design the practical exercises, train assessors and organise validation in real work situations. Discover our AI training for companies, see which AI training to choose for your company or request a diagnostic and an assessment of AI skills.

Sources and References

Français·English·Español