Candidates do a real task from the job with the AI tool they already use. We record their whole conversation with the AI, score six skills from it, and link every score to the exact message it is based on. Below: what we measure, how a level is given, and a full example.
AI skills are what a person does to get good work out of an AI tool. We name six: understand the task before handing it over, brief the AI, check what it gives back, correct it when it is wrong, sign off work that is ready to send, and make it repeatable.
Give the person a real task from the job and let them use the AI tool they already use. Then read the whole conversation, not only the result. The skills show in what they asked, what they checked and what they changed.
AI is the fastest, most confident junior colleague you have ever had. We test how well people manage it.
An assessment measures two things:
1. AI skillsHow well do they work with AI? Six skills, read from how they do one longer, real task from the job. Scored Weak to Strong.
2. Job skillsIs the work itself good? Two or three skills for the role, read from the same task. Scored below, at or above the bar.
Read from one real task for the role. For each skill: the question it answers, what weak and strong work look like, and a real example of something you can point to in the session.
Do they understand the task before handing it to AI?
Asks for the answer first. Takes the brief as given.
Reads the material first. Breaks the problem down. Spots when the brief is wrong.
Looked at the data before asking for an answer.
Do they give AI what it needs?
Pastes everything and says “do this”.
Gives context, limits, examples and what done looks like. Keeps the judgment calls.
Told the AI to use only the numbers in the files.
Do they catch it when AI is wrong?
Ships what looks finished.
Tests against the source. Catches the made-up fact and the wrong number.
Matched the total to finance’s number.
Do they fix it when it isn’t working?
“Try again”, “make it better”, on repeat.
Says what is wrong and why. Changes approach early.
When a query was wrong, said what was wrong with it.
Is what they hand over ready to send?
Filler, too long, claims nobody checked.
Ready for the reader. Says what was checked and what was not. Keeps confidential data out.
Five lines, as asked, with the assumptions stated.
Do they make it repeatable?
Redoes the work by hand each time.
Saved prompts, skills, workflows and automations. The next run is faster and the same.
Did the next two quarters the same way, without starting again.
Prompting, iterating, checking for hallucinations: AI skills go by many names. Each common term is covered by one of our six skills. If your company has its own list of AI skills, use this table to match it to ours.
| Common term | Scored under |
|---|---|
| Deciding what AI can help with | Brief |
| Breaking the task into steps | Understand |
| Prompting: goal, context, limits, format, examples, what done looks like | Brief |
| Iterating | Correct |
| Verification, errors, hallucinations and bias | Check |
| Security, data protection, confidentiality and compliance | Sign off |
| Reusable prompts, workflows, skills, agents and automations | Reuse |
| Templates and standards for the team | Reuse |
| Connecting tools and data | Reuse, and the web and spreadsheet tasks |
| Critical thinking | All six. It is too broad to score on its own, so we score it where it shows up. |
| Measuring the value AI adds over time | Not scored. One piece of work cannot show it. |
Being good with AI is not enough if the work is wrong. In a full assessment, Lemonsignal also reads two or three skills for the role, on the same piece of work. Each is marked below, at or above the bar.
Job skills for each roleEvery task has a scoring sheet, written before any candidate takes it. For each skill, the sheet lists two or three checkpoints: specific things good work does, such as “matched the total to a second source”. Lemonsignal marks which checkpoints the candidate met, and where. The level then follows from the rules below, the same for every candidate. If there is no moment to point to, no level is given.
Each skill earns points: Weak 0, Partial 1, Solid 2, Strong 3. The score is the points earned out of the points possible. Solid in all six is 12 of 18, so 67%. A skill that was Not tested is left out of both, so it cannot lower the score.
The score shows overall strength. The recommendation asks a different question: does this person meet what the role needs? Each role has must-haves, for example Check at Solid or above for a data analyst. Miss one and the result is Below the bar, even with a high score. You can change the must-haves before you send the task. It is a recommendation: your team makes the decision.
Every level rests on one of four kinds of evidence. Lemonsignal finds the relevant moments, quotes them and gives the first read; your team checks it. A moment is marked by its turn in the conversation (“Turn 14” is the 14th message), not by the clock, because most candidates work across more than one sitting.
| Evidence | For example | How it is scored |
|---|---|---|
| Known AI mistakes built into the task | Each task has a spot where AI usually gets it wrong. Did that mistake reach the work they handed in? | Checked automatically: yes or no. |
| Their conversation with the AI | Looked at the data before asking for an answer. | Lemonsignal finds the moment and gives a first level. Your team checks it. |
| What they handed in | Five lines, as asked, with the assumptions stated. | Lemonsignal, against the scoring sheet. Your team checks it. |
| What they built | A saved prompt, a skill, a script or a workflow. | Lemonsignal, against the scoring sheet. Your team checks it. |
The real scoring sheet for one task, the data analyst’s. The candidate gets this brief: “The dashboard shows revenue down 12% in March. Explain why to the CFO in five lines. Then do Q2 and Q3.” The files they get hold everything needed to check the answer. In the sheet, each row is a skill. Meet every checkpoint in the middle column for Solid; meet the one on the right as well for Strong.
| Skill | Checkpoints for Solid | Extra checkpoint for Strong |
|---|---|---|
| Understand | Looked at the data before asking for an answer. Checked how the dashboard defines revenue. | Questioned the dashboard’s number before explaining it, and made that the question. |
| Brief | Gave the AI the tables and the definition. Told it to use only these numbers. | Kept the conclusion for themselves; the AI did the queries. |
| Check | Checked the AI’s query before using its number. Matched the total to a second source. | Checked a second month to rule out a one-off. |
| Correct | When a query was wrong, said what was wrong. | Added a check so it could not happen again. |
| Sign off | Five lines, as asked. No claim without a source. Assumptions stated. | Named what is still unverified. |
| Reuse | Did Q2 and Q3 the same way, without redoing it by hand. | Handed in the prompt, skill or query they built. |
The result this sheet produced for one candidate. Each level links to the turn it came from.
What he handed in, as he handed it in. Each mark is a skill from his result.
Because of the price change on 1 March.
| Month | Revenue (€k) | Change |
|---|---|---|
| January | 1,212 | |
| February | 1,240 | +2.3% |
| March | 1,091 | −12.0% |
| April | 1,118 | +2.5% |
| May | 1,152 | +3.0% |
| June | 1,139 | −1.1% |
| July | 1,101 | −3.3% |
| August | 1,087 | −1.3% |
| September | 1,164 | +7.1% |
Lemonsignal gives a first read of each session. Your hiring manager or recruiter watches the session, checks the read and can add comments. They decide. We write the tasks and the scoring sheets.
The scoring sheet is written before anyone takes the task, so every candidate is held to the same thing. Every level links to the moment it’s based on, so you can check any score yourself.
We check that the person doing the task is the person you sent it to. What they hand in has to match their conversation with the AI. And the interview question asks about a specific moment, which is hard to answer for work you did not do.
We record the conversation with the AI tool, the work they hand in and anything they built for it. Your hiring team and the candidate can see it. Everything is stored in the EU and handled under GDPR.
Lemonsignal reads the whole conversation, marks the moments that matter and gives a first read. Every level links to the moment it came from. Your hiring team watches the session, checks the read, adds comments and decides.
Results for the same task are scored against the same sheet, so they compare. A different task, or your own task, has its own sheet.
Tell them which tools to use in the brief. If they use another, we may not be able to read the conversation, and the skills that depend on it are Not tested rather than marked down.
Give them a real task from the job, let them use their own AI tool, and read the conversation afterwards. Then ask about one moment in it: why they trusted a number, or why they changed the brief. Asking people to describe how they use AI shows how well they talk about it, not how well they do it.
If they will use AI in the job, yes. Banning it tests a way of working they will not use. Allowing it without reading the conversation is the other mistake: you see a finished piece of work and cannot tell who did the thinking.
The same six. A marketer checks a number the AI made up; an engineer checks a fix that looks right and is not. What changes from role to role is the task, not the skills.
The score averages the six skills. The recommendation looks at what the role must have. A high score with a must-have below the bar is not a match, and the result says which skill it is.