Pipeline
Literacy, Fluency, Prompt Engineering: Only One Is a Hiring Bar
AI literacy, AI fluency and prompt engineering are three different things. Literacy is knowing what a model is and how it fails, which is roughly an afternoon of onboarding. Prompt engineering is technique for getting useful output, which is a fortnight of deliberate practice. Fluency is the judgment layer above both: what to hand over, and whether to trust what came back. Hire for fluency. Train literacy and prompt technique, and stop writing them into requirements as though they selected anyone.
The takeA ladder that ends in a course can only sell what a course teaches, and that is the two rungs a competent hire absorbs in two weeks. That is a market rather than a scandal: literacy and prompt technique are teachable, packageable and easy to certify, and judgment is none of those things. A requirement naming the first two can be satisfied by Friday. The bar the training market cannot reach is the only one that does any selecting, and everything below it is a purchase order.
Where Olive fits
Open a role and see what the work shows
Olive is priced per attempt rather than per seat, and one attempt returns six evidenced findings about how a candidate worked through a role-grounded assignment with an AI assistant. Ten attempts a month are free, so a team can see what the judgment rung looks like in practice before writing it into a requirement.
Rank your shortlistWhat does each term actually mean?
Three terms, three different objects. AI literacy is knowledge about the technology: what a model is, what it does with a prompt, and the characteristic ways it goes wrong. Prompt engineering is technique applied to one exchange: framing, examples, constraints, iteration. AI fluency is the judgment sitting on top of both, about which work should go to a model at all and whether what came back can be trusted.
The two frameworks named here are not written for the same reader. The AILit Framework, published jointly by the European Commission and the OECD, organises AI literacy into four domains: Engage with AI, Create with AI, Manage AI and Shape AI 1. It was written for primary and secondary education, for teachers, school leaders and parents, and it carries no test, no scores and no workplace proficiency bar. Lifting it into a job description imports a definition of what a schoolchild should be able to do.
The AI Fluency Framework, written by Rick Dakan and Joseph Feller and turned into courses with Anthropic, defines fluency as the ability to work effectively, efficiently, ethically and safely with AI, and names four competencies: Delegation, Description, Discernment and Diligence 2. Two of those, Discernment and Diligence, are about assessing what came back and vouching for it once it goes out. That is the part a resume cannot show you and an unstructured conversation rarely reaches.
Prompt engineering has no framework problem, because it is a craft rather than a construct. It is the set of habits that produce a usable first draft: state the goal, supply the context, show an example, say what the output is for. It transfers across models with some friction, and it is learned by doing it badly a few dozen times.
Which of the three can you install after the hire?
Literacy and prompt technique, comfortably. An adult who has never used a model reaches working literacy in an afternoon of guided use, and reasonable prompt technique inside a fortnight of real tasks. Judgment about what to delegate and what to verify is the one that resists installation, because it rests on knowing a domain well enough to notice when a fluent answer is wrong.
More technique, applied to work the person could not evaluate, moved the answer further from correct. In a pre-registered field experiment with 758 Boston Consulting Group consultants, one task was chosen deliberately to sit outside what the model could do well. Consultants using GPT-4 were 19 percentage points less likely to reach the correct answer than the control group, and the subgroup given a prompt-engineering overview did worse than the subgroup given none: 24 points below the control against 13 3. One task, one sitting, a 2023 model, so the size of the gap is local to that experiment. A hiring decision inherits the direction and not the number.
Technique earns its keep on work someone can already judge, so the order matters when a req opens. Hire against discernment and the two weeks you spend on tools and prompt habits compound. Hire against tool familiarity and you have bought those two weeks while skipping the thing they were supposed to rest on. The same argument written as what a posting should name is in an AI skill is a decision, not a tool you have open.
Why a certificate does not settle it
Because a certificate reports an exam result, and an exam reaches literacy and prompt technique only. The two named below say on their own pages what they cover: vocabulary, and one vendor's product line. Neither puts a candidate in front of real work with an assistant and records what they refused to accept, because an exam has nowhere to put that.
Read the exam guides rather than the badge names. Microsoft's Azure AI Fundamentals, the credential most often listed as an entry-level AI qualification, states that the candidate needs knowledge of Python coding syntax and programming techniques and familiarity with Azure resources, and the majority of its weighting sits on implementing inside one platform 4. Google Cloud's Generative AI Leader runs the other way: open to anyone in any job role, with or without hands-on technical experience, no prerequisites, ninety minutes of multiple choice 5. Both are honest about what they are. Neither is a proxy for whether someone catches a confident wrong answer.
Training budgets are still worth having. A certificate belongs in a process as a note that onboarding will be shorter, and it settles nothing about the judgment rung. A candidate holding a foundational credential and no habit of checking outputs is precisely the profile the trainable rungs were meant to be layered onto.
When a posting turns a certificate into a hard requirement, it has quietly moved the bar down to the rung the training market sells. That still filters, in the way any arbitrary threshold filters. It selects for people who bought a course.
Write the requirement against the rung you cannot train
Name the judgment, not the vocabulary and not the tool. A usable requirement describes a decision someone can be observed making or failing to make: what gets handed to a model, what gets checked before the work goes out, what stays with a person whatever the deadline. Literacy and prompt technique move to the onboarding plan, where they cost a fortnight and no candidate is excluded for lacking them.
Each of the three terms has one correct home, and putting a term in the wrong home is the entire error:
- Selection criteria. The judgment rung, and nothing else. Decides what to hand over, checks what came back against something outside the conversation, keeps the call that is theirs to make.
- Onboarding, week one. Literacy. What the tools do, how they fail, which of them the company pays for, and the company's own rules about what may go into a prompt.
- Onboarding, weeks two and three. Prompt technique, learned on real tasks with someone reading the output before it goes anywhere. Nobody needs a course for this.
The first row is the only one that has to be defensible, because it is the only one that excludes anybody. A requirement there needs a plain sentence a manager could describe someone failing, and it needs a legal read, which is what how to write an AI-skills requirement that is not legally vague covers.
Then decide, before the req goes live, how that first row gets checked. A requirement nobody assesses is decoration, and discernment shows itself in work with a wrong answer buried in it. An interview question about prompt habits reaches something else. What AI fluency means on a job description, and how to test for it picks up there.
Common questions
Is prompt engineering still worth naming as a skill?
As a trainable habit, yes. As a selection criterion, it has thinned out. The technique that mattered in 2023 was partly compensation for models that needed careful handling, and current models need less of it. What survives is unglamorous: say what the output is for, supply the context the model cannot see, show one example. That is a fortnight of practice for someone who already understands the work, which makes it an onboarding item rather than a filter.
Does anyone define AI fluency the same way twice?
No, and the disagreement is worth knowing before you quote a framework. The AI Fluency Framework, written by two academics and used in Anthropic's courses, names four competencies: Delegation, Description, Discernment and Diligence. The European Commission and OECD framework covers AI literacy in four different domains and was written for schools. Neither publishes levels, scores or norms, so neither can be cited as a proficiency bar. Take the vocabulary if it is useful and write your own observable behaviors underneath it.
Should a posting ask for years of AI experience?
It is a weak proxy and it excludes people for the wrong reason. Widely used assistants are only a few years old, so any figure above two or three years describes when someone started rather than how well they work. Someone with six months and a verification habit will outperform someone with three years of accepting first drafts. Ask for the behavior instead, and let the length of exposure be whatever it is.
What belongs in the onboarding plan if you hire for judgment?
The two trainable rungs, plus the parts that are specific to your company. A short guided session on what the tools do and where they fail. Two weeks of real tasks with someone reviewing the output. Your own rules on what data may go into a prompt and which work has to be checked before it leaves. That is a plan measured in days, which is exactly why none of it belongs in the selection criteria.
Can one interview round separate fluency from technique?
Not reliably on its own. A conversation captures how someone describes checking a claim, which is a different thing from watching them check one, and candidates who prepare well describe it beautifully. What separates the two is work with an error in it: a task where the assistant produces something plausible and wrong, and the only visible difference between candidates is whether it survived. That has to be built into the round rather than asked about in it.
References
- 1. Frameworks - AILit Framework ailiteracyframework.org Supports the four domains of AI literacy and the point that this framework was written for school education, not for a workplace proficiency bar.
- 2. Framework for AI Fluency ringling.libguides.com Supports the definition of AI fluency and the four competencies, including Discernment and Diligence as the judgment layer.
- 3. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality (Working Paper 24-013) mitsloan.mit.edu Supports the claim that prompt technique applied to a task the person cannot evaluate did not help, and on one out-of-frontier task went the other way.
- 4. Exam AI-901: Microsoft Azure AI Fundamentals learn.microsoft.com Supports the claim that a widely listed entry-level AI credential expects Python knowledge and weights most of its content on one vendor platform.
- 5. Generative AI Leader certification cloud.google.com Supports the claim that a deliberately non-technical AI credential is a short multiple-choice exam with no prerequisites and no practical component.
5 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.