How to design a capability programme with AI

Define capability as observable work

Start with the business outcome and the tasks that contribute to it. For each audience segment, describe the conditions, available tools, constraints, expected output, and quality standard. Avoid labels such as “advanced AI user” unless they are translated into observable behaviors.

The US Office of Personnel Management describes needs assessment as identifying the gap between current and desired performance and considering causes before selecting training. The CDC Quality Training Standards emphasize needs assessment, learning objectives, learner engagement, and evaluation. These sources support the design discipline used here; they do not validate a Two Prune programme or benchmark.

Use a capability map:

| Audience | Real task | Observable behavior | Evidence | |---|---|---|---| | Analysts | Produce an evidence-backed option brief | Separates sourced facts, assumptions, and recommendations | Brief, source map, revision record | | Managers | Review AI-supported work | Challenges unsupported claims and approves a bounded next step | Review notes and decision rationale | | Programme owners | Monitor recurring failure patterns | Converts evidence into targeted support changes | Cohort error analysis and change log |

Diagnose before prescribing content

Collect work samples, observation notes, system constraints, interviews, and outcome data. Distinguish a knowledge gap from missing access, unclear policy, weak workflow design, insufficient practice, or absent feedback. Training will not repair every cause.

A common failure is responding to “our teams need prompt training” with a broad prompt library. If the observed issue is that analysts cite inaccessible sources and reviewers cannot reconstruct claims, the capability need is evidence navigation and verification within the real workflow. The intervention may combine a short model, authentic practice, a source-map aid, reviewer coaching, and system changes.

AI may cluster error descriptions and propose hypotheses, but retain the original observations and have domain owners confirm the diagnosis. Do not infer individual ability from sparse activity traces.

Design practice, support, and evidence together

This synthetic programme fragment targets evidence-backed option briefs:

| Component | Design | Evidence produced | |---|---|---| | Model | Ten-minute worked example distinguishing fact, inference, and choice | Annotated reference brief | | Practice 1 | Repair a brief with missing citations | Revised brief and source map | | Practice 2 | Handle conflicting sources under time pressure | Decision note with uncertainty | | Feedback | Reviewer uses task-specific criteria | Criterion notes and revision priorities | | Workplace support | Reusable evidence table and escalation cue | Completed aid in live-like task | | Follow-up | New scenario four weeks later | Transfer sample and error comparison |

The practice resembles the target task and generates evidence at each stage. Completion data alone would show attendance, not capability.

Set measurement and iteration rules

Define baseline evidence, practice checks, transfer evidence, and programme-level review. Use the same essential performance criteria across comparable samples, while documenting scenario differences. Separate participant evidence from programme evidence: one person’s result should not determine whether the whole design works.

Specify in advance what would trigger a change. For example, if many participants frame the decision well but repeatedly lose source provenance, add practice at the evidence-linking step. If performance is sound in practice but fails in the workplace, investigate access, incentives, workload, and manager support before adding lessons.

Verify the programme

Before launch, confirm:

The final artifact is the capability programme map plus diagnostic evidence, practice designs, support plan, measurement schedule, and iteration log.

• each component maps to a diagnosed performance gap;
• practice reproduces the important conditions and constraints;
• assessment requires an authentic output;
• feedback criteria describe observable work;
• workplace aids and manager support have owners;
• transfer evidence is scheduled;
• privacy and access boundaries are documented; and
• AI-generated materials are checked against approved sources and task reality.

Sources

CDC Quality Training Standards

• OPM Training and Development: Planning and Evaluating
• OPM Instructional Systems Series 1750

Sources: [1] [2] [3]