The Kirkpatrick Evaluation Coach skill for Claude

The problem the skill solves is named in the file itself and will be familiar to anyone who has ever organised a training course. A satisfaction survey tells you how people felt. Level 3 and 4 tell you whether anything actually changed. Most evaluations get stuck at level 1, leaving you with enthusiastic participants and a board asking what it delivered. This skill carries the measurement plan through to the figures that board actually looks at, without pretending that this proves the training was the cause.

It is written for training courses, onboarding, compliance programmes and AI adoption programmes. No prior knowledge of Kirkpatrick is needed: the levels, the design logic and the pitfalls all sit inside the file. It is part of the skill library NL that we make freely available through our AI and automation service, with no account and no follow-up sales email. New to skills? What are Claude skills NL explains first exactly what you are downloading.

One thing you should know upfront, and the file is strict about it: this is a measurement framework, not an instructional design and not a scientific effectiveness study. The measurement plan makes the effect plausible and discussable. Without a control group, level 4 shows correlation and not cause, and according to the skill that sentence belongs in your reporting too.

The measurement is only thought up afterwards. By then the training is already designed, the participants are already chosen and the moment for a baseline measurement has passed. The New World Kirkpatrick Model reverses that: you start at level 4, the desired business result, and work back to what people need to do, know and experience to get there. The skill follows that order literally: first the question of which result the training should affect, and only then does it build level 1.

There is no nameable behaviour. Without a concrete description of what people do differently after the training, there is nothing to observe. That is why the skill asks that question first, and why it does not proceed without an answer. The file says it plainly: no nameable behaviour, no measurement plan.

Indicators are feelings, not behaviour. Shows engagement is not an indicator, because two line managers will see something different in it. Shares a method with the team every week is, because it either happened or it did not. It refuses to deliver observation schedules without observable behaviour, and that difference is exactly where most behaviour measurement comes unstuck.

And level 4 gets either skipped or overstated. Skipped because it is hard, or a KPI rise gets attributed entirely to the training. Neither happens here, and that is a deliberate choice. It links three to five existing KPIs with a data source and measurement frequency, and names in the same breath which other factors move the same KPI.

In diff form, with indicators from the skill's example. On the left what usually ends up in an observation form, on the right what you can actually tick off.

Next comes the theory the skill rests on, written out in the file itself. Donald Kirkpatrick described his four evaluation levels in 1959 in a series of articles in the Journal of the American Society of Training Directors, drawn from his doctoral dissertation. In 1994 he brought the model together in Evaluating Training Programs: The Four Levels. His son James and daughter-in-law Wendy updated it in 2016 as the New World Kirkpatrick Model, with more emphasis on designing the measurement upfront and on support after the training.

The core of the file is a list of eight fixed steps. First it asks targeted questions: the training goal, the target group and group size, duration and format, and above all the desired behaviour change on the shop floor. It works back from level 4 and looks for three to five existing KPIs. It builds level 1 as a short survey on relevance, applicability and engagement, with a target score that you set. Level 2 becomes a set of test criteria for a pre- and post-test that match the desired behaviour.

It builds level 3 as an observation schedule for day 30, 60 and 90, with three to five observable indicators per measurement point and a form to fill in. Level 4 becomes the link between those indicators and the KPIs, with measurement frequency, data source and the caveat about which other factors move alongside it. It states the resources needed honestly: a survey tool, a spreadsheet and an observation form are enough. And it delivers a timeline from day 0 to day 90, with the owner of the measurement at each point.

Before anything gets built, the skill runs through a mandatory input checklist: the training goal in one sentence and the desired behaviour change, the target group and number of participants plus whether their line managers are willing to help with observation, duration and format, the existing business KPIs with current values if you have them, the available tools, and the timeframe in which the organisation wants to see results plus who needs to approve the measurement plan. It only asks about what is missing.

The output follows a fixed order of six parts: goal and behaviour change with the assumptions you need to confirm, then the four levels each with their measurement point and owner, and finally the timeline with every measurement point laid out in a row. Want to learn to write files like this yourself? Writing SKILL.md NL explains the structure.

The model comes from Donald Kirkpatrick. He described the four levels in 1959 in a series of articles in the Journal of the American Society of Training Directors, drawn from his doctoral dissertation, and brought it together in 1994 in Evaluating Training Programs: The Four Levels. That is still the skeleton of every measurement plan the skill produces.

The design logic comes from James and Wendy Kirkpatrick. Their New World Kirkpatrick Model from 2016 places emphasis on two things that were underexposed in the original version: designing the measurement upfront instead of thinking it up afterwards, and support after the training. That is where the reversed order in FIG.02 comes from.

The fifth level comes from Jack Phillips. Return on Investment in Training and Performance Improvement Programs from 1997 added ROI on top of level 4, with the accompanying caveats. The skill names that source, but it does not calculate an ROI and does not invent a single percentage. If you want to work out the amounts behind that kind of calculation, you are better off with the Pricing Strategy Coach skill NL, which deals with exactly that value question.

The criticism comes from Alliger and Janak. Their article Kirkpatrick's Levels of Training Criteria: Thirty Years Later from 1989 in Personnel Psychology is about the assumed causality between the levels: the idea that satisfied participants also learn, and that whoever learns also applies it, turns out not to be a given. That source is in the file and explains why the skill never presents level 1 as a predictor of level 2.

Even without installing the skill you can improve your own evaluation with these principles: start with the result you want to see, write down behaviour you can tick off, agree an owner per measurement point, and put in your reporting whatever else was putting pressure on that KPI.

After that, using it is simple: describe your training goal and, above all, what people need to do differently afterwards. If you get stuck unpacking the zip or finding the right folder, Installing Claude skills NL walks through it step by step. And if you want the broader explanation of working with AI first, you will find it in the knowledge base.

There are also situations where you are better off leaving it aside, and the file names them itself. Designing the training is one of them: this is a measurement framework, not an instructional design. What happens in the training and how you build up the material is work for a different kind of help, for example the Bloom Taxonomy Coach skill, which writes learning objectives and test questions per thinking level. Statistical proof is the second limit: the measurement plan makes the effect plausible and discussable, it is not a scientific effectiveness study with a control group. And the third is assessing people: the observation scores are there to evaluate the training, not to fill out a personnel file.

One more honest limit the skill names itself: with small groups and short timeframes, level 4 is often hard to prove. That does not mean you skip it, but that the reporting honestly states what you can and cannot see. If you are working towards a decision on whether to continue or stop, that distinction matters more than the figure itself.

The hardest question in this whole framework is the first one: what do people demonstrably do differently after the training. Anyone with a clear answer to that has the measurement plan half done, and usually a better training too. Want your team to get going with AI and be able to show straight away what that changes? That is exactly the combination we deliver in our AI training NL. And if you want to talk further about what else AI can take on in your organisation, from loose skills to full automation, we simply do that in a conversation.