12 Meta-Reflection Techniques to Make AI Smarter

With the current leap in AI represented by Claude Code and OpenClaw, we are increasingly aware of the remarkable intelligence of AI. However, you can leverage some meta-reflection techniques to make it even smarter.

Perhaps no discipline values the concept of "meta" more than cognitive science. In cognitive science, there is the study of the cognition of cognition - meta-cognition; the study of the memory of memory - meta-memory; and the study of the language of language - meta-language. Such studies are extremely common in cognitive science. What is meta-reflection? Similarly, it is the reflection on reflection.

These are some meta-reflection techniques I often use in practice for your reference.

Group 1: General

Technique 1: Meta-reflection

Instruction example: Please use "deep thinking" (invoke sequential-thinking mcp) to conduct a round of meta-reflection. Check if your implementation is correct, if it achieves our goals, if there are any major logical errors or omissions. In particular, be aware that when achieving new goals, you may disrupt old functions.

This is the technique I use most frequently. As long as you let AI conduct meta-reflection, new problems will surely be discovered. AI is a vector matrix with thousands of dimensions. It is better than humans at making connections and finding pathways that we previously didn't notice. This is its advantage. However, the drawback is that any neural network mechanism inevitably has limitations in information processing. This applies to both humans and large AI models. Because the working mechanism of a neural network is essentially the activation of weights/parameters. So-called activation means it is destined to be imprecise and cannot be precise.

Therefore, the meta-reflection process is extremely necessary for AI. Apart from the built-in meta-reflection process in models like the think class, we also need to actively force AI to conduct meta-reflection at key points. My experience is that for almost any complex task, as long as meta-reflection is carried out, AI will surely surprise me.

Technique 2: Constraint Entities

Instruction example: Please list the entities involved in this round of tasks. Are there no more than 4 entities? If there are more than 4, use TaskCreate to split them into subtasks.

Large AI models still have "bounded rationality". No matter how the context window is designed, the best dialogue mode is still a session that deals with no more than 4 entities. In database terms, it means being constrained within 4 tables.

Once the number exceeds 4 tables, the large model will most likely make mistakes. This has nothing to do with using a more advanced large model. It is due to the inherent flaws brought about by the large model's imitation of the human brain structure. As long as it is a neural network architecture, such defects are inevitable.

Technique 3: Redefine the Problem

Instruction example: Please use "deep thinking" (invoke sequential-thinking mcp) to deeply understand the problem. Does the problem exist? Can the problem be redefined? Can the boundary of the problem be narrowed or converted into a problem that has a mature solution in the industry or another field?

Often, defining the problem is more important than solving it. Redefining the problem and optimizing its boundary actually solves most of the problem. Once, I was struggling with a strategic problem. But when I asked AI to convert it into a timing optimization problem, it became much easier to solve. A problem that has no mature answer in one field may have a solution in another.

Group 2: Intent

Technique 4: Set a Safety Net

Instruction example: Have you considered a safety net strategy?

The so-called safety net means what to do if things go wrong. Like humans, AI is limited by the training time. It often tries to give you an answer in the shortest possible time. This answer is usually from a positive perspective, that is, how to succeed. But it never considers what to do if it fails.

Technique 5: Triple Protection

Instruction example: Have you considered triple protection?

This technique is similar to the safety net strategy, but the protection is more rigorous. It is generally applicable to critical scenarios. As the old saying goes, misfortunes never come singly. In a production environment or a complex project, either there is no failure, or once it fails, if you only set up one line of defense, it is often not enough when a real accident occurs. Three consecutive lines of protection are needed to ensure that everything is okay after a failure.

Technique 6: Self-verification

Instruction example: Before execution, write a verification script for yourself or propose a set of verification criteria. After execution, automatically execute your verification script or use the verification criteria to check for any deficiencies in your execution and continue to optimize.

Before letting AI execute complex tasks, set verification criteria for itself first, and then align with them during the execution process. If there are errors in the verification criteria, continue to optimize. Repeat this cycle. By doing a good job in data monitoring and process design, the automatic optimization of AI will reach a level that will shock everyone. The recently popular ralph-loop method is an excellent example:

https://42plugin.com/plugins?q=ralph-loop

However, this kind of loop method underestimates the complexity of real work. So, I developed my own better dev-loop method, which has worked very well in practice.

Group 3: Implementation

Technique 7: Watchdog

Instruction example: Please use the CronCreate tool to create a watchdog for me. Read the memory file (or a file where you store the work progress) to find the next uncompleted step and execute it immediately. Don't ask me any questions and just execute the next step.

Another similar usage is heartbeat monitoring. The instruction example is as follows: Read the memory file to check if there are any pending reminders or tasks. If there are, execute them immediately without asking me any questions; if not, reply HEARTBEAT_OK.

Why is the watchdog or heartbeat monitoring important? Because it allows us to get rid of the human in the loop and leave the conversation with AI to do other things. Recently, I was live-streaming AI doing work for my classmates. I was running in an outdoor park, and AI was doing the work autonomously at home for 90 minutes straight without any human intervention. It was the watchdog that was working behind the scenes.

Technique 8: Completeness

Instruction example: Have you considered atomicity/transaction consistency/idempotence?

These terms have slight differences. Atomicity emphasizes indivisibility, either do it all or don't do it at all; transaction consistency emphasizes legal states, going from a legal state A to a legal state B without an illegal intermediate state, that is, either fail or succeed, and there cannot be an intermediate state of half failure and half success; idempotence emphasizes harmless repetition, that is, the result of executing once is the same as executing ten times.

But the overall direction is similar, emphasizing that when we take action, we should be more comprehensive: ensure that the operation and result are complete and clean - leave no semi-finished products, generate no dirty data, and don't make mistakes due to repetition.

Technique 9: Align! Align! Align!

Instruction example: Please carefully compare your current implementation with the plan or spec (specification) documents you initially wrote. Is it aligned? Are there any major logical errors or omissions?

The key to making AI work at 10 times the speed is: Align! Align! Align!

Actually, alignment = leveraging the completion ability of the large model. If A is imperfect, B can complete it; if B is imperfect, A can complete B.

If humans are imperfect, AI can complete humans; if AI is imperfect, humans can complete AI.

If Project 1 is imperfect, Project 2 can complete it; if Project 2 is imperfect, Project 1 can complete it.

If Module 1 of a project is imperfect, the relatively more perfect Module 2 can complete it; if Module 2 is imperfect, the relatively more perfect Module 1 can complete it.

Repeat this cycle. This is the real era of "hybrid intelligence"! This is working at 10 times the speed!

This is the key to this round of large models!

So, finding the relatively perfect point is the starting point or key to designing the entire "10 times speed workflow".

In any work, there is a "relatively perfect" point. If there isn't, then "create something out of nothing". Refer to the perfection of others to complete your own imperfection.

All techniques are based on this: finding the relatively perfect point is the starting point for designing the entire workflow.

And how to judge "perfection" is what I mentioned in my New Year's speech "The Intentional Revolution: Towards the Era of Hybrid Intelligence" as the "confirmation cost" and "confirmation coefficient".

Group 4: Feedback

Technique 10: Speed Test

Instruction example: Have you considered more extreme situations? If you want to increase the speed or output quality of your previous execution by 10 times, 100 times, or 1000 times, will your execution still be affected? If so, how can you improve your previous execution?

Often, it's difficult to get good answers when thinking within the same order of magnitude. But once you start thinking about how to achieve 10 times, 100 times, or 1000 times the output or speed increase, good answers are more likely to be obtained.

For program development, a common approach is concurrency and asynchrony. So, you also need to remind AI: Have you considered a high-performance version? For example, changing to concurrent or asynchronous execution? If you have, have you considered the strategy when concurrent or asynchronous execution fails? That is, set a safety net strategy, referring to Technique 4.

Technique 11: Similarity Scan

Instruction example: Please think carefully again. Are there any similar errors? Especially variations of the previous error?

Force AI to draw inferences from one instance and scan for similar problems. This is a very common instruction in my practice. Often, limited by AI's context window, if you don't point it out actively, it won't actively scan for similar problems. For example, if you fix the collection bug in the homework area, it won't care about the same collection bug in the discussion area or the course content area.

Technique 12: Accumulation Flywheel

Instruction example: Please append the newly discovered patterns in this round of work to the memory/dev-loop-patterns.md file.

The pattern understanding ability of large AI models far exceeds that of humans. It is especially good at discovering some patterns that we haven't noticed. By accumulating these patterns, future similar work will be much easier.

This technique can often set indicators and focus on improving key indicators, such as continuously improving a certain indicator in large model training. Combining with the aforementioned techniques will produce extremely amazing results. This is how I wrote a Skill in an AI autonomous work system I created:

Newly discovered patterns → Append to memory/dev-loop-patterns.md

Key indicators → Append to memory/dev-loop-metrics.md

Summary

Students who are familiar with me may recognize the idea behind the above techniques. Isn't this the "Action Loop" I talked about in the "Meta-knowledge" series of lectures?

Yes, that's exactly the answer. The Action Loop includes four elements: the intention of action, the implementation of action, the feedback of action, and the context in which the acting subject is located. To improve human's ability to take action, we should work around these elements. Similarly, to improve AI's output, we still need to work around these elements. Only, AI's context becomes the context (Context).

Currently, mainstream cutting-edge large AI models generally receive reflection-based training during post-training, such as the Chain of Thought (CoT). But having AI conduct meta-reflection in an appropriate context will achieve better-than-expected results. After all, during the training of large AI models, it's impossible to cover all the contexts in your work.

You can arrange these techniques into Skills, Commands, Agents, etc. That's what I do. I've formed a powerful AI autonomous work system, which can easily enable AI to work autonomously for a long time. Recently, for fun, I held five live streams of AI working autonomously. Without any intervention from me, in real complex projects, AI can easily work autonomously for 90 minutes to 3 hours. This is the key to this round of AI leap: without any intervention, how to let AI independently complete a large or difficult production-level task?

Humans eat and exercise, while AI works diligently in the live stream. This is the ideal future for humans.

This article is from the WeChat official account "Mind Toolbox". The author is Yang Zhiping. It is published by 36Kr with permission.