The Prompt Engineering Skill Matrix: What Good, Better and Best Prompting Looks Like by Job Family

Why "good with AI" keeps slipping out of performance reviews

A finance manager asks two analysts to summarise the same variance report. One returns a clean three-paragraph note with the reconciling items flagged. The other returns a confident paragraph that quietly invents a figure. Both spent forty minutes on the task. In a review, both would probably be described as "comfortable with AI" — a phrase that tells you nothing you can act on.

That is the practical problem with AI capability inside a department. Most teams now have some AI usage, but very few have a shared definition of what competent usage looks like. Prompts are treated as personal style rather than as a work skill with observable behaviours. The result: uneven quality, unspoken risk, and performance conversations that drift into impressions.

A prompt engineering skills matrix fixes that by turning prompting into three named levels — aware, applied, advanced — each defined by what a person can be seen doing, not by how enthusiastic they are about tools.

The three levels, defined by observable behaviour

Keep each level to a single sentence a manager could verify from a sample of work.

Level 1 — Aware

Can use AI for low-risk, clearly bounded tasks and recognises when output needs checking.

• Writes a simple instruction with a stated task and audience.
• Spots an obviously wrong fact, tone problem or hallucinated reference.
• Knows which data must never be pasted into a public tool.
• Escalates when unsure rather than guessing.

Level 2 — Applied

Can structure a prompt with context, constraints and an output format, and improve the result through deliberate iteration.

• Supplies role, context and constraints before asking for output.
• Requests a defined structure — table, bullet list, email with subject line.
• Reviews drafts against a source document rather than trusting the first answer.
• Keeps a personal set of reusable prompts for recurring tasks.

Level 3 — Advanced

Can design a repeatable prompting workflow that others reuse, and can judge when an off-the-shelf tool is the wrong instrument.

• Chains multiple steps: extract, restructure, critique, then rewrite.
• Builds and documents prompt templates for a team, with examples and guardrails.
• Tests prompts against edge cases and records where they fail.
• Identifies workflows that need custom AI solutions or automation rather than a better prompt.

Notice that none of these levels is about clever wording. Level 3 is about repeatability and judgement — the two things that make prompting a department capability rather than an individual hobby.

Which tool for which job: a comparison managers can act on

The matrix only works if you also accept that tool choice is part of the skill. These four tools behave differently on the same task, and the right answer depends on the work, not on preference.

Preparation task	Strongest fit	Prompting skill it tests	Watch-out
Summarising a policy document for staff	ChatGPT or Copilot in a licensed tenant	Level 2 — constraints on length, audience and citation	Confidential drafts must stay inside the approved tenant
Reformatting a data extract into a table	Copilot inside Excel	Level 2 — specifying column order and formats	Verify totals against the source file
Drafting a customer reply from a knowledge base	Gemini or a custom internal bot	Level 3 — retrieving the right source first	Never let the model invent policy detail
Building a monthly report template	ChatGPT with a saved prompt chain	Level 3 — repeatability and documentation	Version the template so it does not drift

Two practical rules fall out of this table. First, tool licences do not create capability — a team on the best seat in the market with only Level 1 behaviour will still produce checkable-by-hand output. Second, where a workflow requires retrieving from many internal sources, prompting harder is the wrong lever; that is usually an implementation question.

Role by role: what each level looks like in real work

Finance

• Aware: drafts a variance commentary outline, then reconciles every number manually.
• Applied: produces a structured variance note with flagged reconciling items, each traced to the ledger extract.
• Advanced: builds a reusable prompt set for monthly close commentary and documents the review step for the financial controller.

HR

• Aware: drafts a job description from a rough brief and edits it manually.
• Applied: generates screening question sets aligned to a competency framework, with a consistency check across candidates.
• Advanced: designs a policy Q&A prompt workflow with an approved source list, and routes anything outside it to a human.

Operations

• Aware: summarises a supplier email thread.
• Applied: converts an SOP into a step-by-step checklist with defined escalation points.
• Advanced: maps where prompting stops helping and where automation, or a custom AI workflow, should take over.

Marketing

• Aware: asks for headline options and picks one.
• Applied: briefs the model with audience, tone, proof points and a banned-claims list before drafting.
• Advanced: maintains a brand-voice prompt library with regression examples so output stays consistent across the team.

Customer service

• Aware: drafts a reply to a routine enquiry.
• Applied: produces a response with the policy reference included and a clear escalation line.
• Advanced: defines the prompt and escalation logic behind a customer enquiry chatbot, including what it must refuse to answer.

A note on tool-specific training

The same levels apply to a Microsoft-heavy environment; the habits differ more than the levels do. AIHQ's guidance on AI classes for professionals Malaysia covers how feature-based Copilot training differs from habit-based training.

Turning the matrix into evidence in a performance conversation

A level becomes useful only when it produces something you can look at. Four artefacts do most of the work:

1. A task log. Three real tasks per quarter, with the prompt used and the final output. Two minutes to record, hard to argue with later.
2. A verification habit. Every AI-assisted deliverable carries a visible check — a source link, a reconciled figure, a named reviewer.
3. A shared prompt file. Each team keeps five to ten prompts that are known to work. New joiners start at Level 2 faster.
4. A named failure. Each person can describe one time the tool was wrong and what they did about it. This is the single best signal of Level 2 judgement.

When these exist, the review sentence changes from "comfortable with AI" to "consistently operates at Level 2 on reporting tasks; ready for Level 3 template ownership next quarter." That sentence is coachable, and it is fair across roles because the behaviours are role-relevant, not tool-specific.

Calibrating expectations across five roles

Departments rarely need everyone at Level 3. A workable target profile looks like this:

• Analyst and executive roles: Level 2 as the standard, Level 1 as the floor.
• Managers and team leads: Level 2 plus the ability to review someone else's prompt output.
• Specialists in regulated work: Level 2 with a documented verification step; Level 3 is optional.
• AI champions, one or two per department: Level 3, owning the shared prompt file.
• Anyone handling personal or client data: Level 1 knowledge of guardrails is non-negotiable before any tool access.

Two constraints keep the matrix honest. Prompting is one capability among several — it does not substitute for role-based judgement, governance or leadership alignment. And a matrix with no worked examples becomes a poster on a wall, which is why organisations often need a structured AI innovation bootcamp to generate real prompts from real workflows before levels mean anything.

A 90-day way to put this in place

• Weeks 1–2: pick two job families and write three observable behaviours per level. Use the artefacts above as your evidence sources.
• Weeks 3–6: run a short practical session per job family. Everyone leaves with one task log entry and one verified output.
• Weeks 7–10: have champions draft the shared prompt files. Managers review them once, together, to calibrate what Level 2 looks like.
• Weeks 11–13: apply the matrix in one round of one-to-ones, then adjust the wording where managers disagreed.

If the calibration step exposes a wider gap — inconsistent team practice, no agreed guardrails, no review habit — that is a capability-building question rather than a vocabulary problem. Structured role-based AI training is usually the shortest route, and AIHQ's work with over 9,000 professionals across corporate, public sector and regulated environments has been built around exactly that sequencing.

For teams moving from individual prompting to shared workflows, the next decision is whether the constraint is skill or system. A consultation on your AI adoption needs is a reasonable place to settle it.