Prompting training for AI-enabled knowledge work

Teach prompts as part of a work system

Start with the decision and deliverable. Identify the audience, inputs, source hierarchy, constraints, uncertainty, material errors, and verification standard. Show how a prompt expresses those requirements, but keep responsibility for evidence and judgment with the person doing the work.

Separate reusable moves from model-specific syntax. Useful moves include defining the task, supplying relevant context, requesting a structured output, limiting sources, asking for alternatives, exposing assumptions, testing a counterexample, and requiring traceability. Their value depends on the model, configuration, tools, and task.

The NIST GenAI evaluation programme studies generators, detectors, and prompting strategies across modalities and emphasizes measurement of capabilities and limitations. It supports an evaluation mindset, not a claim that one prompting pattern reliably works everywhere.

Sources: [1]

Build the prompting practice deck

Create several cards around one representative workflow. Card one asks for an initial plan from a bounded brief. Card two adds an authoritative source set. Card three introduces conflicting evidence. Card four requires tool use or calculation. Card five presents a plausible but unsupported output for critique. Card six asks for a final deliverable and an evidence note.

For every card, record the target behavior, task materials, prompt or instruction, model configuration, output, checks, revision, unresolved issue, reviewer anchor, and lesson. Change one important condition at a time so the participant can explain what altered the result.

Design accessibility into the deck. Provide equivalent access to source materials, allow assistive technology, and avoid grading typing speed or verbosity when those are not part of the construct. Mark any card that fails to create a fair evidence opportunity.

The UK government's 2026 employer guide to AI upskilling recommends practical and contextualized learning connected to real roles and workflows. Its findings support task-based training but do not validate a particular prompt curriculum.

Sources: [2]

Worked artifact: a synthetic supplier brief

Imagine a synthetic procurement exercise that asks a participant to prepare a supplier-renewal brief from an agreement, performance table, service email, and policy excerpt. The first card produces a generic summary because the instruction lacks audience and decision context.

The participant revises the prompt to identify the renewal committee, required recommendation, controlling agreement, evidence table, and unresolved questions. The assistant then cites a service promise from the email as if it changed the contract. On the conflict card, the participant instructs the model to separate contractual obligations, operational reports, and proposals, then verifies the classification against original sources.

The final artifact includes the initial and revised instructions, output excerpts, source links, correction, unresolved issue, and explanation of why the prompt changed. The lesson is not a magic phrase; it is that source authority and verification must remain visible in the work.

The scenario is synthetic. Performance in the deck does not establish durable transfer, future job performance, or a customer outcome.

Verify prompting capability without counting prompts

Review whether the participant framed the task, selected evidence, delegated appropriately, noticed failure, verified material claims, and improved the deliverable. Prompt length, number of turns, and use of fashionable terms are weak substitutes for these behaviors.

Retest with a different case and, where relevant, a changed model or interface. Record missing opportunities and external checks that telemetry cannot see. Two Prune can provide controlled assessment evidence about AI orchestration and verification for human interpretation; it does not prescribe a universal prompt formula or make employment decisions.