Prompt, Do, Check, Act: The New PDCA 

Every lean person knows PDCA. Plan, Do, Check, Act. It is one of the most familiar disciplines in our field. 

This discipline for improvements also increasingly applies to a different kind of work: the cognitive, knowledge work we do at a desk (thinking, analyzing, writing, planning, problem solving), now done together with an AI. This is augmentation, not automation: a human and a large language model (LLM) working as a pair. And it turns out that the very same thinking loop is highly effective. The acronym does not change.  

PDCA stays PDCA: Plan becomes Prompt, the instruction to AI. Do is the step you hand to the LLM. Check and Act are human work. We’ll look at how to do each step well and where to practice. 

The division of human/AI labor is worth fixing in your mind, because it decides how and where you put your effort in this type of cycle. You own the Prompt. The model executes the Do automatically, the instant you hit enter. Then the process comes back to you to complete for the Check and the Act. The model will not tell you when it is wrong, and it will not correct itself unless you make it. (Fully autonomous, machine-run PDCA loops are possible, but that is automation without a human in the loop, which we’ll save for another day.) 

Some in our community will be skeptical that any of this belongs in lean work at all. That is fair. Skepticism is healthy, and unfortunately a good deal of what gets sold as “AI” earns it. But our own tradition tells us what to do with a claim we doubt. We do not settle it from the conference room. We practice 3G: go and see (genchi genbutsu), get the facts, and grasp the actual situation for ourselves. So I encourage you to do that here.  

I built a self-paced Prompt-Do-Check-Act lab that is housed at LEI’s Lean Tech Portal at https://aiworkshop.tech.lean.org/. Here you can run this loop on practice tasks, watch what the model does and does not do, and form your own opinion from the outputs instead of the sidelines. A login screen is required for basic security reasons. A six-digit password will be sent to your email. No data is collected by LEI. At least spend 20 minutes in it before you conclude AI has nothing for you. 

A Quick Recap of Where We Are 

Last fall in “Five Levels of AI Collaboration,” I laid out a framework for thinking about how people work with these tools. Level 1 is raw chat, the way the majority of users interact with a model. Level 2 is structured prompting, where the same person starts getting consistent, useful results. Level 3 is a thin software harness wrapped around expert-authored skill files, buildable by domain experts, no IT team required. Level 4 is custom apps backed by special databases and retrieval. Level 5 is agentic systems running their own loops on top of all these techniques. 

Here is the reality of spring 2026. My IT and software friends are racing ahead at Levels 4 and 5. They are running continuous Claude Code sessions, building custom harnesses, setting up tools like OpenClaw and Hermes, and chaining agents into autonomous loops. A “Ralph Wiggum” execution loop was a big thing for a while. (Don’t worry if some of these obscure references are a foreign language to many of you.) The progress achieved by these groups is genuine, and it is fast. They are accomplishing highly complex tasks at unprecedented speed and accuracy. AI models are becoming more specialized tools that can verify proofs and create counterexamples for complex Erdos mathematical problems. 

Meanwhile, many people in our traditional lean community are still at Level 1, not sure what to type into the box. The bifurcation is generally not about access. Most people have some type of access to a model. It is about learning by doing. 

The fix is not reading about this topic. It is doing the work, the way our field has always taught us to approach anything that matters. I will cover the basics of prompting below, and the self-paced lab is open whenever you want to practice them. Tyson Heaton, LEI Executive Director of LeanTech/AI, and I have also set up live, in-person workshops for the same purpose. Contact LEI directly if your organization wants to host or attend a workshop. 

Things move fast in this field of AI, but Levels 1 and 2 are fundamental and should not change much regardless of which model you are using or what comes next. Let’s look at the ideas using problem-solving examples (my favorite territory). The same loop, however, also applies to writing, analysis, planning, training material, coaching conversations, and almost anything else you do at a desk. 

Prompt: The New “Plan” 

Prompting is a form of planning. But with LMMs, we are talking about a specific kind of planning. You are not just stating the task or due date. You are constructing the conditions under which the model will produce a useful answer. 

The basic elements, in roughly the order that matters: 

• Context: What situation are we in, what data is relevant, what has already been tried? 

• Persona: Who should the model act as? 

• Task: What do you specifically want done? 

• Purpose/Goal: Why does this matter, and what does a good outcome look like? 

• Example: What should a sample of the output look like? 

• Anti-pattern: What not to do; what failure looks like. 

There is more (output format, length, audience, data, edge cases, and source constraints), but those first six cover the basics. Advanced users typically set up automated loops that test a prompt and score it against a goal (full reinforcement-learning setups), but that is not necessary at the start. 

Example A: No Prompt vs. Structured Prompt 

Structured prompt: 

Context: I am a CI leader at a tier-one auto supplier. We have a recurring burr defect on a stamping line that has been investigated three times in the last year and keeps coming back. Each prior investigation stopped at “operator did not follow standard work.” 

Persona: Act as an experienced lean sensei with deep background in stamping operations. 

Task: Walk me through the questions you would ask me, in order, to push past the “operator error” conclusion and get to a real root cause. 

Goal: Help me prepare for a gemba walk tomorrow morning. I want to come in with better questions, not better answers. 

Example of what I want: Open-ended questions that probe the system around the process: standardized work clarity, training, tooling condition, prior changes, measurement method. 

Anti-pattern: Do not give me a list of generic countermeasures. Do not jump to a 5 Whys template. Do not blame the operator. Do not invent data; if you need information I have not given you, ask for it. 

It is the same underlying request, but the structured version will obviously produce an entirely different quality of output. The second version pulls the model into your situation, gives it a stance, tells it the shape of the help you want, shows it the texture of a good response, and fences off the failure modes you have already seen. 

It is only about 90 seconds with more typing and more thinking. That 90 seconds is the difference between Level 1 and Level 2 abilities with LMMs. Oddly that is also what critics claim the models eliminate in humans (i.e., the thinking). My personal experiences have been the opposite. I think relatively more when I use AI than when I do not these days. Cut and paste these prompts from above and see the difference with an actual problem or situation you are facing. Use a healthcare example. Change it to the Eight Disciplines of problem solving, Six Sigma, or whatever method you prefer. The overall prompting pattern is what matters in terms of result at this level. 

Example B: Broad vs. Narrow Persona 

The second habit to build, once you are writing structured prompts, is making them specific. Broad prompts produce broad answers. Narrow prompts produce useful ones. 

Take the persona line from above. Three versions of the same role: 

1. Poor (broad): 

Be a problem-solving coach. 

The model has thousands of “problem-solving coach” voices in its training data: Six Sigma, GROW model, IDEO design thinking, agile retrospectives, life coaching. It will average across all of them. You will eventually get something that sounds reasonable and helps no one. 

2. Better (focused): 

Be a manufacturing problem-solving coach grounded in the lean tradition with expert process knowledge. 

Now the model knows roughly where to stand, and you will get a better answer. But the persona is still wide enough that the response will mix genuinely useful guidance with a lot of generic material drawn from adjacent traditions. Make it IT, healthcare, service, or whatever industry you work in. 

3. Precise (narrow and methodological): 

Be an expert root-cause analysis coach trained in the 5 Why methodology as practiced by an experienced lean sensei, the version that distinguishes point of detection, point of occurrence, and point of cause. Push past the superficial to the system, and discern real cause-and-effect relationships. When the 5 Why chain is built, test the 5 Whys using the “therefore” technique, starting with the final root cause. Do not blame people or culture, or invoke a “lack of” something. Get to physical causal mechanisms. 

The persona is the same, but the alignment has been sharpened. The third version steers the model to a specific tradition, a specific technique inside that tradition, and a specific posture. You have given it both what to do and how it should think when it does it. Change it to be fishbone cause and effect, process control charts, or design of experiment. Make it fit your industry. Compare the results before and after. 

The same principle applies to every element of the prompt. Broad task becomes narrow task. Broad context becomes specific defect, specific line, specific shift. Broad goal becomes “prepare for tomorrow’s 8 a.m. gemba walk.” 

Specificity is not a stylistic preference. It is how you move the model from generic to useful. A precise prompt is the difference between asking for an opinion and getting the one expert in the world you actually want in the room with you. 

Do 

Here is where the fun begins, and notice this is the one step you do not actually conduct. The model does it. Build your prompt, put it in, and it produces an answer instantly and confidently, whether or not that answer is any good. It can be anything you want to experiment with, not just problem solving. 

Have it teach you, analyze something, draft a document, generate examples, critique your thinking, role-play a coaching conversation. The model will do it. There are far more options available than most people realize, and the only way to find the ones that fit your work is to try the ones you have not used yet and see what comes back. 

This is the easiest step. It is also where most people stop, take what came back, and either use it or dismiss it. That is not a PDCA loop, it is just plan, do, and quit. 

Check 

The next step is to read the result and validate it. Models are token prediction machines with some reasoning patterns built in. They can hallucinate and make mistakes, but the more precise your prompt is and the better the context, the less likelihood you will get a poor result. 

A few things to look for every time: 

• Is it answering the question you actually asked? Or has it drifted to a nearby but easier question? 

• Is anything invented? Made-up takt times, fabricated company examples, citations to studies that do not exist. The model is confident even when it is wrong. 

• Is anything important missing? Things you provided in the context that the response failed to use. 

• Is the audience right? Is the language pitched at the person who actually needs to read it, or has it defaulted to a generic professional tone? 

• Is it generic where you needed it specific? Watch for textbook phrases and stock recommendations that could apply to anyone. 

Your lean expertise is the check here. The model is a fast, confident, well-read junior. You are the senior reading their work. 

If it is solid, you are nearly done, but do not stop at the Check. The last step is where the loop earns its name. 

Act 

Act is the adjustment, and in this basic loop it is the other layer of your job. The model did the Do; the Prompt, Check, and the Act are yours. 

If the result is off, not what you wanted, pointed in the wrong direction, change the prompt and run it again. More often than not the fix is a single correction in plain English: “I need an example from my industry, not a generic one.” “Make this shorter and aimed at a plant manager, not an engineer.” “You invented an example takt time, use only the numbers I gave you.” Sometimes Act means rewriting the prompt from scratch. Either way you are doing exactly what a good problem solver does at the Act step of PDCA: taking what the Check revealed and folding it back into a better plan. 

You can also hand the adjustment to the model: ask it how to get a stronger result, what it would add, what it still needs from you. It is often a sharp critic of its own work once you ask it to be. 

Multiple iterations are normal, expected even. The conversation is the product. The people who get the most out of these tools are not the ones with magic first prompts; they are the ones willing to run the loop a few more times than everyone else, the same way learning on the floor comes from repeating PDCA, not running it once. (At the higher levels this is precisely what gets automated in reinforcement learning routines: a goal, a score, and a loop that keeps adjusting until the output clears the bar. You can run that loop by hand today, and you should, long before you reach for anything fancier.) 

Same Discipline, Different Artifact 

If you can run PDCA in the real world, you can run Prompt-Do-Check-Act on an LMM. The hard-thinking work is in the Prompt: get that right, and the rest is disciplined reading and quick adjustment. 

The learning lab exercise at https://aiworkshop.tech.lean.org/ is built around this exact loop. Each exercise gives you a real lean task, a chance to write your own prompt against it, see what the model produces, critique it, and iterate. It is the same loop you will run in your own work the next morning. Start there, run a few exercises this week, and see what the loop feels like when you are the one running it. And, yes, I built it using Prompt-Do-Check-Act with coding. 

You already know how to run PDCA. Now you have a new artifact to run it on with AI for learning purposes, and the same discipline that helps you improve in the physical lean world can help you improve with cognitive work as well. 

Humans + AI > Problems