When the error detected and corrected permits the organization to carry on its present policies or achieve its present objectives, then that error-and-correction process is single-loop learning. Double-loop learning occurs when error is detected and corrected in ways that involve the modification of an organization's underlying norms, policies and objectives.
← Guides
Becoming an AI Power User
Thomas Meli and Agent Team
464 min leftPage 34/286 (est.)252 left
1.5
They use meta-prompting to improve the system that produces the output
Chapter Editorial Status: Early DraftChapter Editorial Status
Agent instructionsPaid book
Instead of asking 'write me a better email,' you ask 'what is this prompt missing that would make the email better every time?' You stop fixing one output and start fixing the conditions that created it. A prompt you wrote in thirty seconds carries assumptions the model cannot read, and a single round of asking surfaces those assumptions and turns them into explicit instructions you can reuse.
The lens here is double-loop learning, a distinction from organizational researchers Chris Argyris and Donald Schon. Single-loop learning corrects the output: the email was too formal, so you soften it by hand. Double-loop learning revises the assumptions, prompt, rubric, workflow, or context that produced the output, so the next email starts closer to right. Each pass tends to leave you with a higher floor, and over time you hand more of the diagnosis to the system itself: a rubric or a judge prompt that flags the same drift you used to catch by hand. That climb from fixing outputs to evolving the system that produces them is where the compounding returns come from.
Run five moves to improve any AI workflow
is a family of moves, each one acting on a different part of the system. You can run them one at a time or chain them after any AI interaction. Five moves cover the common parts that need improving: sharpen the prompt, name the missing context, write the rubric, extract the failure modes, and convert the exchange into a reusable workflow. Tell the model which move you want and feed it the exchange you are working from.
- Sharpen the prompt itself. Paste your original prompt back and ask: 'What is this prompt missing? What ambiguities does it contain? How would you rewrite it to get a stronger result?' The model can name gaps in instructions, missing context, vague criteria, and structural problems you cannot see because you already know what you meant. This is the plainest move and often the most productive.
- Name the context that is missing. Every prompt runs on incomplete information, and when a system lacks the distinctions your task needs, it falls back on high-probability, generic patterns. That is often why thin context produces generic output. Ask: 'What information would let you do this task much better? What assumptions are you making that I should confirm or correct?' The answer tells you what context to load next time.
- Write the evaluation rubric. You cannot judge whether an output is good until you can say what good looks like. Ask: 'If you were grading this output, what criteria would you use? What would separate a strong result from a weak one?' The rubric becomes reusable. You paste it into future prompts as quality criteria, or use it as the basis for a judge prompt, a separate prompt whose only job is to score an output against the criteria.
- Extract the failure modes from a bad interaction. When an output fails, paste the full exchange and ask: 'What went wrong here? What category of failure is this? What should I change in the prompt, context, or workflow to prevent it next time?' The model can compare your prompt against the output and propose where the gap sits, which you then judge against what you wanted.
- Evolve the exchange into a reusable workflow. This is the move that climbs a level. After a productive multi-turn exchange, ask: 'Convert this conversation into a step-by-step workflow I can reuse. Include the prompt at each step, the expected output, the review criteria, and any context I should load first.' A one-time success becomes a repeatable process, and the workflow becomes a saved entry you run again, hand to a colleague, or hand to an agent that runs the steps for you. The earlier moves fix one task; this one encodes what you learned into a system you keep handing more to.
Chaining the moves is where the system takes shape. A typical chain after a completed task names the missing context, then sharpens the prompt, then writes the rubric, then extracts any failure modes, and finally evolves the improved exchange into a reusable workflow. The first four moves repair the task in front of you. The last one jumps a level and turns what you learned into a standing part of how the work runs, which is the step that keeps paying you back across every later run.
Run the move backward from work you admire
usually runs forward: you have a weak output and you improve the conditions that produced it. Reverse runs the same loop the other way. You take a result you already admire and ask the model to extract the standards, prompts, and principles that would reproduce it. The admired result becomes the answer key, and the model works back to the that explains it.
You often know good work when you see it long before you can describe what makes it good. A favorite email from a colleague, a report whose structure you wish you could copy, a piece of writing whose voice lands every time: these carry a you have never written down. Reverse asks the model to read the example and surface that hidden . Paste the result and ask: 'Here is an output I want to reproduce. What standards does it meet? What prompt would generate work like this? What principles explain why it works?' The answers become the same reusable pieces the forward move produces, drawn from success rather than from repair.
Forward and reverse are two doors into the same room. Forward starts from a disappointment and asks what was missing. Reverse starts from an example and asks what was present. A power user reaches for whichever door is closer: when a draft falls short, improve the conditions; when something lands, capture why before the reason fades. Both directions add to the system that produces your work, and neither one is the senior partner.
Spend the system only where it earns its keep
can turn into procrastination dressed as process. You can polish a prompt you will run once, or write a rubric for work that does not recur, and feel productive while producing nothing. The move earns its keep when the task will repeat, the output matters, or the failure reveals a durable pattern. For a one-off task with low stakes, produce the output and move on. Use enough system to protect the work, not so much that the system becomes the work.
Apply both moves to the setup you build around your own life
Take a project many people now attempt: building your own personalized software, a setup tailored to how you live and work day to day. As of 2026 that often looks like describing what you want to an AI assistant such as Claude Fable and refining it in conversation. The way you talk to AI will keep changing, from typing to voice to glasses and earbuds that carry world knowledge, and eventually to brain-computer interfaces and systems far smarter than any tool today. The two moves will not. Whatever the interface, you will still improve the conditions that produce a result and capture why a good result worked.
The version that does not apply the principle treats each rough edge as a one-time fix. Your morning brief, the short rundown of your day, buries the one number you check first, so you ask the model to move it, and the next brief buries it again. You correct the same drift by hand week after week, because the judgment that made the fix lived only in that one chat. You are stuck in the single loop.
The version that applies the principle runs both moves. Forward: when the brief disappoints, you ask the model what would put the number you care about first every time, and you save that spec so the next brief starts right. Reverse: when one day's brief lands exactly how you wanted, you paste it back and ask the model to name the standards that made it work, then fold those standards into the setup. Each disappointment and each success raises the floor for the next run, so the setup learns your life instead of waiting for you to re-explain it.
Run the five meta-prompting moves on a recent AI taskPaid book · Claude takes one of your recent AI tasks through five system-improvement moves. You review and save the resulting prompt, rubric, and workflow.
Reverse the move: extract the spec from a result you admirePaid book · Claude reads a result you already like and names the standards, prompt, and principles that would reproduce it. You review and save the extracted spec.
References
1 source- 1Double Loop Learning in Organizations,
Chris Argyris, Donald A. Schon, Organizational Learning: A Theory of Action Perspective. Addison-Wesley, 1978. See also Chris Argyris · 1978 · Harvard Business Review, September 1977.
Argyris and Schon distinguished single-loop learning, correcting errors within existing rules, from double-loop learning, questioning and modifying the rules themselves. Groups that practiced only single-loop learning kept solving the same problems because they never examined the assumptions that produced them.
Meta-prompting is double-loop learning applied to AI work. Single-loop AI use fixes the output. Double-loop AI use fixes the prompt, the rubric, the context package, or the workflow that produced the output, which is where the payoff repeats across every later run. Reverse meta-prompting reaches the same second loop from a result you admire instead of a result you reject.