The best discussion podcast content on Prompt Engineering seen to date.

AI Summary

summarize

AI Tips Engineering In a roundtable format, several Anthropic experts shared their understanding and practical experience of cue engineering from different perspectives: research, consumer and corporate.

The article details the definition of cue engineering, its importance, and how to become a good cue engineer.

The core idea is that cue engineering is more than simple text input, but a process that requires constant iteration, experimentation, and a deeper understanding of the psychology of modelsThe

It involves how to effectively communicate with AI models and integrate them into the larger system.

The similarities between cue engineering and programming are also explored, as well as the different emphases in different application scenarios (e.g., research, business, and everyday conversation).

emphasizes Clear communication, ability to iterate and meticulous observation of model outputs It's the key to prompting the project.

Experts also discussed various challenges encountered in practical applications and shared valuable experiences and tips, such as how to deal with edge cases, how to utilize feedback from the models themselves to improve cues, and how to distinguish between different types of models.

In a nutshell.This article provides readers with a comprehensive and insightful guide to prompt engineering and a look at the future direction of theThe

key point

• Cue engineering is a process that maximizes the potential of a model by communicating clearly with it and iterating.
• At the heart of engineering is experimentation and trial and error, the ability to go back and try different approaches from scratch.
• Prompts are not just text, but a programming approach integrated into the overall system that requires consideration of data sources, latency, and system design.
• Good cueing engineers need to be able to communicate clearly, iterate, and analyze model output in depth.
• An understanding of the model's "mind" is critical, taking into account how the model interprets instructions.

Innovative insights

• Think of the written word as code, and consider writing a good essay as important as writing code.
• Emphasize the importance of reading the output of the model, similar to "looking at the data" in machine learning.
• Offer to use the model to help optimize the cue, or even have the model point out its own errors.
• It is argued that explaining the task directly to the model, rather than pretending to be a character, is more effective in many cases.
• The future of cue engineering will be more focused on getting information from the user, and models will play the role of guides.

Key Citations and Translations

Original article 1: I think the engineering part comes from the trial and error. So one really nice thing about talking to a model that's not like talking to a person, is you have this restart button.

Translation: I think the "engineering" part comes from trial and error. Well, one of the very different things about talking to a model and talking to a person is that you have a reset button. This giant button takes you back to square one and starts from scratch.

Reason for citing: This quote eloquently points out the meaning of "engineering" in cue engineering, and emphasizes the importance of continuous experimentation and iteration to improve cues, which is key to differentiating cue engineering from other forms of communication.

Original 2: I think of prompts as the way that you program models a little bit, that makes it too complicated. 'Cause I think Zack is generally right that it's just But if you think about it a little bit as programming a model, you have to think about where data comes from, what data you have access to... (paragraph 7)

Translation: I think hints are kind of like a way to program the model, although it's a bit complicated to say so. Because I think Zack's right that clear communication is what's important. But if you think of prompts as programming a model, you have to think about where the data is coming from and what data you have access to.

Reason for citing: This passage links cueing to programming, emphasizing that cueing is not just about language, but requires systematic thinking that includes factors such as data, latency, and systems integration.

So I think that I can trust the model if I look at 100 outputs of it and it's really consistent. And I know that I've constructed those to basically figure out all of the edge cases and all of the weird things that the model might do, strange inputs, et cetera. And I know that I've constructed those to basically figure out all of the edge cases and all of the weird things that the model might do, strange inputs, et cetera. I trust that probably more than a much more loosely constructed set of several thousand.

Translation: So if I look at 100 outputs of a model, and they're very consistent, and I know that I've constructed those outputs to figure out all the edge cases and weird things that the model might be doing and weird inputs and so on, I'm going to trust that model, and it's probably a lot more trustworthy than one that's constructed loosely with thousands of outputs.

Reason for citing: This passage emphasizes the value of small datasets of high quality over large datasets of low quality. In cue engineering, the focus should be on adequate consideration of edge cases rather than blindly pursuing a large number of samples.

Reading Notes

[Hints on the nature of engineering]: Iteration and experimentation

• Cue engineering is the process of optimizing model output through trial and error and iteration.
• It emphasizes managing and tracking experiments like programming.
• Clear communication is the foundation of cue engineering.

#promptengineering #iteration #experimentation

[Qualities of a cue engineer]: Understanding and observation

• Clear communication skills and a willingness to iterate are required.
• Requires a deep understanding of the model and the ability to learn from the model output.
• Need to take into account the diversity of real-world usage scenarios and user inputsThe

#communication #understanding #observation

[Model Interaction]: Trust and Questioning

• Models should not be trusted blindly and need to be tested and validated for reliability on an ongoing basis.
• Models can be used to self-diagnose errors and suggest improvementsThe
• Models can help users better understand tasks and be prompted.

#trust #feedback #collaboration

• Cue Engineering Process::

Start --> Write initial prompt --> Model output --> Analyze results --> Modify prompt --> Loop until satisfied

Similar to a cyclical process, tips are constantly improved to achieve the desired output.

• Model Evaluation::

| Output Consistency | Edge Case Coverage | Input Diversity |
| ------------| ------------| ------------|
| High | High | High | High |

• The higher the three key dimensions, consistency, edge case coverage and input diversity, the better.

Tips for the future of engineering

User Requirements --> Model Understanding --> Model Questions --> User Feedback --> Optimization Tips --> Final Outputs

• In future cue engineering, models will be more proactively involved in the process of user requirements understanding and cue optimization.

Questions and answers on core issues

1. What is a prompt project?

Prompt Engineering is a technique for guiding large language models (LLMs) to perform specific tasks by designing and optimizing prompts (Prompts). It aims to enable the model to produce the desired output through clear and precise communication. The following is a detailed explanation of Prompts Engineering:

Definition of cue

• directives: A prompt is an instruction provided to the model by the user to get the model to perform a specific task. It can be a simple sentence or a complex description containing multiple steps.
• programs: Hints can also be viewed as a program, written in natural language, to guide the model in performing a task.
• communication method: Essentially, cues are a way of communicating with the model, similar to communicating with a person, and need to be clear and unambiguous.

Cue the core elements of the project

• Clear communication: Cue engineering emphasizes clear communication so that users can accurately express their needs and so that the model can understand the specific requirements of the task.
• iterative process: Cue engineering is an iterative process that involves continually trying, modifying, and optimizing cues to improve the performance of a model. It is similar to the development and debugging process in software engineering.
  • test: Observe the model's response by trying different cues and adjusting the cues based on the results.
  • send back information: Analyze the output of the model, identify errors, and correct accordingly.
  • circulate: Repeat the experimentation and feedback until the desired result is achieved.
• systems thinking: Cue engineering is not just about writing individual cues, but also about considering how cues are integrated into the overall system. This requires consideration of the source of the data, how it is processed, and the role of the model in the system.
• Making Sense of Models: Cueing engineers need to understand how models work and their limitations in order to better design cues. This includes understanding how the model handles different types of inputs and how to guide the model's reasoning process.
• Problem solving skills: Tip Engineers need to systematically consider all possible scenarios and provide solutions for them as if they were engineering problems.
  • error in forecasting: Predict how the model is likely to go wrong and devise appropriate hints to deal with these errors.
  • Dealing with edge cases: Consider how the model reacts when it encounters unusual inputs or errors.
• version control: Treat hints like code, including version control, tracking experiments, etc.
• Reading Model Output: Read the output of the model carefully to understand its reasoning process, not just to see if the results are correct.
• theoretical thinking: When understanding a model, you need to consider how the model might understand your instructions at a theoretical level, not just write them based on your own understanding.

Purpose of the reminder project

• Realizing the potential of models: The purpose of cue engineering is to fully utilize the potential of a model so that it can perform tasks beyond its original design capabilities.
• Optimizing Model Performance: Improve the model's performance on specific tasks with well-designed cues.
• Guiding Model Behavior: Guide the model's behavior through cues that enable the model to produce desired outputs and avoid undesired outputs.

Cue Engineering Challenges

• Difficult to articulate: It is difficult to describe a task accurately in words and remove all assumptions.
• Models don't ask questions.: Models do not ask clarifying questions like humans do, so cueing engineers need to anticipate possible model queries themselves and give the appropriate answers in the cue.
• The Difficulty of Finding the Perfect Tip: Finding the perfect tip can be a challenging task because there is always the possibility that a better tip exists.

Cue Engineering Applications

• Various scenes: Cue engineering can be applied in a variety of scenarios, including research, enterprise applications, and consumer applications.
• Different types of tasks: Prompt projects can be used for different types of tasks, including text generation, information extraction, question answering and code generation.
• Integration into systems: Prompt engineering is not just about writing individual prompts, but also about integrating prompts into the overall system.

Tips for the future of engineering

• Model-assisted Hints: In the future, the model will be able to help users write better prompts, including asking questions, offering suggestions and automatically generating prompts.
• human-machine collaboration: In the future cue engineering may shift to a human-computer collaboration model, where the model asks questions based on the user's goals and guides the user in writing more effective cues.
• From guidance to counseling: As models become smarter, cue engineering may shift from a guidance model to an advisory model, where the model will reverse cue the user based on the user's goals.

Prompt engineering is a technique that requires creativity, logical thinking, and systems thinking. It's not just about writing a good cue, it's about understanding the model, designing experiments, iterative optimization, and problem solving. Cue engineers need to experiment and learn like engineers in order to realize the full potential of a model.

2. What are the qualities of a good cue engineer?

• Clear communication skills: Good prompting engineers are able to articulate ideas clearly, understand tasks clearly, and describe concepts accurately. This includes the ability to construct instructions in a way that is easy for the model to understand.
• Iterative capabilities: They are willing to continually iterate and tweak the prompts and think about where the model may be misinterpreting them. This iterative process involves analyzing the model's response, identifying errors, and making corrections.
• Test Edge Cases: They will actively think about the less common situations in which the cue might go wrong, such as when the input is null or not in the expected format. This involves testing a variety of exceptions to ensure that the model works properly in different situations.
• Understanding Model Outputs: Good cueing engineers pay close attention to the output of the model, not just the results. They will delve into the thinking process of the model and try to understand its reasoning.
• theoretical thinking: They are able to pull from their own understanding and systematically break down all of the information needed to accomplish a task. They are able to clearly communicate the necessary information in a way that the model can understand.
• empathy: They can put themselves in the shoes of the model and understand how the model perceives their instructions. They also need to take into account the needs of the user and understand how the user interacts with the model.
• experimental mindset: They discover the boundaries of the model through constant experimentation and trial and error. They are not afraid of failure and learn from it, deepening their understanding of the model by pushing the limits of its capabilities...
• Improvements using models: They will improve the prompts not only through their own efforts, but also by using the model itself. For example, they will ask the model to point out ambiguities in the instructions or ask the model to suggest changes. They will try to get the model to explain its errors and improve the instructions.
• Trust but verify: They are cautious about the model's capabilities and ensure its reliability through repeated testing. They verify the output of the model through extensive testing rather than blindly trusting the model.
• scrutiny: They will read the prompts and the model's output carefully and analyze the details. They will learn how the model thinks by analyzing the details of the model's output.
• Not obsessed with perfection: Instead of striving for the perfect prompt, they will keep trying and learning from their mistakes. They will recognize that prompts are an iterative process, not a one-time task.
• Treating text as code: They can understand the meaning of text as code and understand that hints also require version control, experiment tracking, etc.
• Ability to think from different perspectives: Good cueing engineers can think about things from different perspectives, for example, putting themselves in the shoes of the model as well as considering the needs of the actual user.
• Ability to create new concepts: They will define new concepts as needed and clarify them by working with models.
• Ability to externalize ideas: They can express their ideas clearly and can translate complex concepts in their brains into instructions that models can understand.

Good cueing engineers are not only clear communicators and iterative, but also empathetic, able to put themselves in the shoes of the model, and able to constantly experiment and learn, discovering the boundaries of the model by testing edge cases. They will also use the model itself to improve cues and learn from the details of the model output. They need to understand that the process of getting to the cue is iterative rather than striving for perfection, and they need to understand the similarities between text and code. They need to think about it from different perspectives, both in terms of the user experience and in terms of the way the model itself is perceived. Most importantly, they need to be able to articulate their ideas and externalize the concepts in their brains.

3. How to effectively interact with the model?

3.1 Clear communication is at the core

• Accurate expression of needs: Just like communicating with people, you need to clearly express your needs so that the model understands exactly what you are aiming for.Avoid ambiguous instructions, try to be as specific as possible in describing what you expect the model to accomplish.
• Clarify the details of the task: You need to.Strip away all your assumptions., and detail all the information the model needs to know.Don't assume the model knows anything you haven't explicitly told it.The

3.2 Consider the prompt as a program

• Natural language code: A prompt can be thought of as a program written in natural language to guide the model through a task.
• Systems Thinking: Treat prompts like code, including version control, tracking experiments, etc. Consideration needs to be given to how cues fit into the overall system, including data sources, data processing, and the role of models in the system.

3.3 Embrace iteration and experimentation

• Trial and error is the norm: Cue engineering is a trial-and-error process that involves constantly trying different cues and adjusting them based on feedback from the model.
• Restart button: The model has a "restart button" so that you can go back to the beginning at any time and try a different approach from scratch, without being disturbed by previous experiments.
• Frequent iterations: Effective cueing engineering requires frequent interaction with the model for multipleiterate back and forth, rather than expecting perfect results all at once.

3.4 Understanding the Mind of a Model

• Modeling Perspectives: Try to think about your instructions from the model's point of view and understand how the model might understand your requirements. This requires you toPlaying the Modelcharacter that mimics the way it thinks.
• Read the model output: Read the output of the model carefully, focusing not only on whether the results are correct, but also on understanding the model's reasoning process.Learning from Output, understand how the model understands your instructions.
• Explore the modeling error: Don't ignore the mistakes the model made.Ask why the model is wrongand try to understand the cause of the error, and may even ask the model to modify the instructions. Models are sometimes able to point out ambiguities in instructions and give suggestions for improvement.

3.5 Dealing with edge cases

• Prediction error: Anticipate situations where the model may be wrong, and design appropriate cues to handle these errors. Consider how the model reacts when it encounters unusual inputs or errors, for example:
  • Provide options: If the model is not sure what to do with certain inputs, give it an "exit", e.g. let it output the label "not sure".
  • Test for extremes: Try testing your prompts with a variety of extremes (e.g., empty strings, inputs that don't match expectations) to make sure the model works well in a variety of situations.

3.6 Ability to respect models

• Don't underestimate the model: Don't think that the model is stupid and needs to be "coaxed" to work.Be respectful of the model's capabilities, give it enough context and information so that it can understand your goals.
• Give direct information: When you want a model to learn a new technique, you can just give it the relevant paper or document instead of trying to describe it in your own words.
• Avoid oversimplification: Don't deliberately simplify your instructions; trust the model to handle complex tasks.

3.7 Use of modeling aids

• Model generation example: Using the model to generate examples that you can then modify allows you to generate high-quality prompts faster.
• Modeling for interviews: Let the model interview you, extract the information in your head, and turn that information into prompts.

3.8 Don't fetishize the perfect tip

• There is no perfection: Don't fall into the trap of "looking for the perfect tip" and realize that there is always room for improvement.
• Continuous learning: Every interaction with the model is a learning opportunity, and every attempt will allow you to understand the model better.
• Focus on boundary exploration: Try to get the model to do something you don't think it can do, and learn by exploring the boundaries of the model.

3.9 Distinguish between different scenarios

• Research and Enterprise: In a research environment you may be more focused on diversity and exploration, while in a corporate environment you may be more focused on reliability and consistency.
• Human-computer dialog and system applications: In a human-computer conversation, you might do multiple iterations, but in a system application, you need to write prompts that can handle a variety of situations all at once.

3.10 Using Meta Prompts

• Generate prompts for prompts: You can use "meta-hints" to have the model generate the output or query you need. You could just give the model a paper on the hinting technique and have it generate a meta-hint to get other models to perform the technique.

In short, interacting effectively with models requires clear communication, systematic thinking, consistent experimentation, deep understanding, and respect for the model's capabilities. At the same time, effectively using the aid of models can help you iterate and optimize your prompts faster. Remember, there is no such thing as a perfect cue, only continuous learning and improvement.

4. Common misconceptions about Prompt

4.1 Hints are just simple instructions:

• Misunderstanding: People often think of prompts as simple instructions given to a model, like typing keywords into a search engine. They may think that all they need to do is type in some keywords to get the model to complete the task, and overlook the importance of clear and precise communication.
• Facts: In fact.Hints are a complex way of programmingthat need to be treated like code, including version control and tracking experiments. Good cues need to be carefully designed and iterated to ensure that the model understands the task accurately and produces the desired output.

4.2 Tips are static and can be written once:

• Misunderstanding: Some people think that writing a prompt is like writing an essay; it's done and no further revisions are needed.
• Facts: The hint project is aiterative process, which requires constant experimentation, modification and optimization. You need to work with the model several times tointeract back and forth, and improve your cues by reading model output and analyzing errors. Effective cue engineering requiresEmbrace experimentation and feedback, rather than expecting a one-step process.

4.3 Hints need perfect grammar and punctuation:

• Misunderstanding: It is often assumed that in order for a model to be understood, the prompt must use perfect grammar and punctuation.
• Facts: While attention to detail is important, theModels often understand prompts that contain spelling errors or grammatical imperfections. Important.Conceptual clarity, not grammatical perfection. While it's good to fix errors in the final prompt, it's acceptable to be imperfect during iterations.

4.4 Models need to be "coaxed" to work:

• Misunderstanding: Some people believe that the model is stupid and needs to be guided through tasks using tricks or "lies", such as giving the model a false identity or role.
• Facts: Models have a strong understandingYou don't need to "coax" it. You should.Respect model, and give it clear, accurate information directly so that it understands your goals.Describe your task directly, rather than using metaphors or similar tasks to guide the model.

4.5 The heart of cue engineering lies in writing perfect instructions:

• Misunderstanding: Some people think the point of cue engineering is to find the perfect instructions and spend a lot of time figuring out each word.
• Facts: While precise instructions are important, it is even more important that theUnderstanding how the model worksand throughRead the output of the model to learnTheUnderstanding the mindset of the modelas well asHow it handles different inputs, more than the quest for perfect instructions. A good cueing engineer should be able toExtracting signals from the output of the model, and understand its reasoning process, not just whether the result is correct.

4.6 Tip Engineering is just writing:

• Misunderstanding: Some people believe that the core competency of prompt engineering lies in writing skills, believing that good writers naturally make good prompt engineers.
• Facts: While good writing skills are necessary, itNot a core competency for cue engineering. A good prompting engineer needs to haveExperimental spirit, systems thinking, problem solving skillsas well asThe ability to understand the model mindTheIteration and TestingMore than mere writing skills.

4.7 One should avoid giving too much information to the model:

• Misunderstanding: Some worry that giving the model too much information will confuse it, so they try to simplify the instructions and hide the complexity.
• Facts: As the modeling capabilities increase, they are able to handleMore information and contextTheYou should trust the model., give it enough information so that it can better understand your task.

4.8 More example tips are always better:

• Misunderstanding: One might think that providing a large number of examples is the only way to improve model performance.
• Facts: While examples are helpful in guiding the model, too many examples can beLimit creativity and diversity in modeling.. In the research setting.Use illustrative examples rather than specific examplesmay be more effective because it encourages the model to think about the task itself rather than just copying the example.

4.9 Models think and reason like people:

• Misunderstanding: One might think that a model would reason like a human being and understand the "thinking steps" cues.
• Facts: While models can mimic the reasoning process, for example through theChain-of-thought (chain of thought), but it's not necessarily true reasoning. The model just generates text based on the instructions and examples you give it. It's important to understand thatModels and humans think differently, don't over anthropomorphize the model's behavior.

4.10 Role-playing prompts always work:

• Misunderstanding: Some believe that having the model play a specific role (e.g., "You are a teacher") improves its performance.
• Facts: Role-playing prompts may work in some situations, but are not always necessary. Describe directly what you want to accomplishthat is more effective than pretending the model is a different person. As the model's capabilities improve, it may be better to give the task description and context directly, rather than giving it a false identity.

4.11 Once you find a good tip, it will always work:

• Misunderstanding: Some people think that once they find a cue that works, it will work forever and won't need to be tweaked.
• Facts: As modeling capabilities continue to improve, theEffective tips can also be outdated. Some cueing tricks may be trained into the model so that it no longer needs to be explicitly cued. You will need toContinuous learning and adaptation, in response to changes in the model.

Understanding these common misconceptions can help you interact with models more effectively and better utilize cue engineering for a variety of tasks. Hint engineering is not just a simple command input, but a discipline that requires in-depth understanding and practice.

5. Enterprise tips vs. research tips

Enterprise Tipscap (a poem)Research Level TipsThere are significant differences in objectives, methods and focus.

Enterprise Tips

• Emphasis on reliability: In enterprise applications, reliability is critical. The goal of enterprise-level hints is to ensure that the model produces consistent and expected results in a variety of situations. This usually requires a large number of examples and specific guidance to limit the model's degrees of freedom.
• Focus on format:: Enterprise tips are very much focused on the format of the output. For business applications, stability and consistency of the output format is often more important than variety, as it affects the efficiency of the user interface presentation and subsequent data processing.
• Focus on user needs: Enterprise-class prompts need to be highly responsive to the specific needs of the user. This means that prompts need to be able to process a variety of different inputs and generate output that meets the specific needs of the user.
• Systemic thinking:: Enterprise-level cueing often requires cueing to be integrated into a larger system for consideration. This includes considering data sources, latency, and how to integrate the model with other software and processes.
• Lots of testing and iteration: Enterprise tips need to be tested under a variety of inputs and scenarios to ensure a high degree of reliability and stability in real-world applications. This includes testing a variety of edge cases, as well as a variety of possible user inputs.
• Focus on consistency: In enterprise applications, even if the answer is repetitive, it is acceptable as long as it meets expectations. This is different from exploratory goals in a research environment.
• Focus on long-term applications: Enterprise tips are designed to build a system that can be reused many times over. As a result, Enterprise Hints requires more time and effort to ensure that it works reliably.
• Avoiding too much abstraction: Enterprise-level prompts should avoid overly abstract instructions and instead clearly describe the task and the desired output.

Research Level Tips

• Emphasis on diversity and exploration: The goal of research-level prompts is to explore the various capabilities of the model and to discover possible new uses for the model. This usually entails reducing the constraints on the model and encouraging it to explore different outputs and solutions.
• Favors few or no examples: In order not to limit the scope of exploration of the model, research-level prompts typically reduce the number of examples or do not provide specific examples.
• Focus on cognitive tasks: Research-level prompts focus more on cognitive tasks, i.e., how the model understands and solves complex problems.
• Use of illustrative examples: When research-level prompts provide examples, these tend to be illustrative rather than concrete. This means that the examples may be different from the data that the model actually needs to work with, and is intended to help the model understand the nature of the task rather than directly mimicking the examples.
• Trying out new boundaries: The goal of the research-level prompts is to challenge the boundaries of the model's capabilities and discover what the model does well and what it does poorly. This includes attempting tasks that the model is not good at in order to better understand the model's limitations.
• Increased focus on flexible and diversified outputs:: Research-level prompts may focus more on exploring what kinds of outputs a model can generate, even if those outputs are not highly consistent. Research-level prompts focus more on how the model thinks, and the quality and depth of its outputs, rather than just whether the results are correct.
• More exploratory: Research-grade cues are more exploratory and may focus less on consistency or format. Researchers will focus more on how the model reacts in the face of a new situation and how the cue can be used to guide the model in the direction of exploration.

summarize::

• Different objectives: Enterprise-level prompts are designed to solve real-world problems, emphasizing reliability and consistency, while research-level prompts are designed to explore model capabilities, emphasizing variety and innovation.
• Different methods: Enterprise-level prompts typically employ a large number of concrete examples to control the output of the model, while research-level prompts typically employ few or no examples to encourage the model to explore new possibilities.
• Differences in focus: Enterprise-level prompts focus on user requirements and system integration; while research-level prompts focus on cognitive processes and model boundaries.
• Different development and testing cycles:: Enterprise-level hints typically need to be run in production environments for long periods of time and therefore require more rigorous testing and quality control, while research-level hints may have shorter testing and iteration cycles aimed at exploring the various potentials of the model.
• Treating models differently:: Enterprise-level prompts sometimes "accommodate" the model to ensure that it is understood correctly, while research-level prompts tend to "respect" the model's capabilities and give it more autonomy.

The fundamental difference between enterprise-level prompts and research-level prompts is their purpose and focus.

While enterprise-level hints are dedicated to providing users with reliable solutions, research-level hints are dedicated to expanding our understanding of modeling capabilities.

In practice, these two cues may require different approaches and techniques.

6. Tips for the future of engineering

6.1 Models will be better at understanding your intentions, but clarity is still important

• An information theory perspective: In the future, the model will better understand your needs, but you still need to provide enough information to clarify your goals. Even if the model is able to understand beyond your words.Clearly articulating your expectations remains criticalThe
• The importance of clear objectives: No matter how smart the model becomes.Ability to clarify objectives remains central. Even though models can set goals, if you want to use them to solve a problem then you still need to explicitly specify what you want them to do.
• Ongoing communication: Even as models become smarter and better able to understand your intentions, you still need toCommunicate with models, provide feedback, and make adjustmentsThe

6.2 Models will be your cue assistants

• Collaboration with models: In the future, you will be able to collaborate more deeply with the model to determine what needs to be written and what is missing. The model will help youDiscover what you may not have thought ofand provideSuggestions for improving the tipsThe
• Model-assisted generation of hints: You can use the model to generate examples, drafts, and meta-prompts to speed up the prompt development process. For example, you can use the model to generate examples that you then revise, which is much easier than writing the perfect answer from scratch.
• High-bandwidth interactions: In the future, you will be able to have high-bandwidth interactions with the model, such as providing feedback and requesting adjustments from the model. This interaction will be similar to collaborating with a designer, where you provide high-level goals and the model helps you flesh them out.

6.3 Meta-tips will become more important

• Use hints to generate hints: In the future, you may spend more time looking for hints that enable models to generate desired outputs or queries. You will use meta-prompts to get models to perform specific prompting tricks or to generate prompting templates for other models.
• Give modeling learning materials: Instead of writing your own cues, you can give models relevant papers or documents to learn new cueing techniques. Models can read the papers directly and apply the knowledge from them to cue generation.

6.4 Cue engineering will focus on the boundaries of the model

• Ability to explore models: You will continue to explore the boundaries of the model's capabilities and challenge what the model is capable of accomplishing.
• The pursuit of excellence in performance: You will focus on getting the highest level of performance from the model and exploring what the model can barely accomplish.

6.5 The model may, in turn, suggest that you

• The model understands your intentions: When models know more than you do about the context of the task, they may prompt you in turn to clarify your needs. Models may ask questions to help you clarify what you are trying to accomplish and to uncover edge cases that you may have overlooked.
• From Instruction Recipient to Expert Advisor: The model will transform from a simple recipient of instructions to an expert advisor that you can consult with about the details of the task. It's like working with a designer; they'll ask you questions to better understand your needs.
• Modeling Interviews: To better understand your needs, the model can come and interact with you like an interview.

6.6 The future requires greater introspection

• The model understands you: In the future, the model will need to understand your ideas, rather than you trying to understand the model.
• Make yourself visible to the model: You will need to learn how to express your ideas and needs clearly so that the model can understand your intentions.
• Define the concept: Sometimes you need to create new concepts and define what they mean so that the model understands your intentions.

6.7 Cue engineering may become a philosophical practice

• Clarity of Expression: In the future, cue engineering may requireThink and write like a philosopher, using clear, precise language to express complex ideas.
• Writing for the educated common man: You need to write the prompt as if you were writing for the educated layperson so that even someone unfamiliar with the subject can understand your intent.
• Externalize your brain: Good cue engineering requires you to externalize the ideas in your brain and make them understandable to the model.

6.8 Tip engineering skills will transfer to higher level tasks

• From low level missions to high level missions: As the model progresses, you will no longer need to focus on prompts for low-level tasks, but will instead focus on higher-level tasks such as task decomposition and complex reasoning.
• Guided Interaction: Future interactions are likely to be more like guided conversations than models typing text into a console to reach a final outcome.

Hints that the future of engineering may requireStronger collaborative, introspective and expressive skills. You will need to work with models to explore their capabilities and define your needs. In addition.You also need to continue to learn and adapt to changes in the model, rather than looking for a one-and-done solution. While the future of prompt engineering will change, clarity of purpose and articulation will remain at the core.

7. Tips for prompting works

The focus is on howImprove efficiency and effectiveness of communication with models::

7.1 Iteration and experimentation:

• Keep trying: The hint project is aiterative processthat requires constant experimentation, revision and optimization. Don't expect to write the perfect prompt the first time, but be prepared to make multipleinteract back and forthThe
• Learn from your mistakes: When the model is wrong, analyze it carefullyReasons for the error, and improve your prompts accordingly. Every interaction with a model is a learning opportunity.
• Embrace the experiment: Be open to trying different methods to see which works best. The heart of the cue project isExperimentation and feedback, rather than one step at a time.

7.2 Clear and precise communication:

• Clearly articulate the task: expense or outlayClear, concise languageDescribe the task you want the model to accomplish. Avoid vague or ambiguous terms.
• Provide sufficient information: Don't be afraid to give the modelProvide detailed context and background information. Make sure the model understands your goals and the specific requirements of the task.
• Respect the model: Instead of trying to "cajole" the model, it is important toRespect modelof comprehension. Describe your task directly, rather than using metaphors or fictional characters.

7.3 Understand how models work:

• Read the model output: Read the model's output carefully to understand itsmindsetand reasoning processes. Observe how the model handles different inputs and adjust your prompts accordingly.
• Explore the boundaries of the model: Try to get the model to do tasks that you think it might not be able to do toUnderstanding the scope of the model's capabilities. This can help you better understand the limitations of the model.
• Try to play the model: Try to put yourself in the shoes ofThinking in terms of models, understand how it perceives your instructions. This can help you better predict the model's behavior.

7.4Diverse cueing methods:

• Example of use: By providingtypical exampleto guide the model through the task. However, be careful not to over-rely on examples as this may limit the creativity of the model.
• Use meta-tips: Use the prompts toGenerate additional tips, or have the model generate output that meets specific needs. This can help you explore different cueing strategies more efficiently.
• Chain Thinking: lit. let the model beA step-by-step explanation of its reasoning process. This gives you a better understanding of the model's decision-making process and helps improve the model's performance.
• Role Playing: While not always necessary, in some cases it is useful to have the modelPlaying a specific roleIt may help to accomplish the task. However.It's often more effective to express your mission directlyThe

7.5 Advanced Tips:

• Define the concept: In order to communicate your intentions, sometimes you need toDefining new conceptsand explain what they mean.
• Let the model interview you: Have the model interview you in turn toHelps you clear your mindthat extracts the information you need to provide to the model.
• Drawing on Philosophy: Learn from philosophical writing how toArticulate complex ideas, so that the model can understand your intentions.
• Inform the model: Instead of writing your own prompts, you can just give the models the relevant papers or documents and let them learn on their own.

7.6 Caution.

• Don't focus too much on grammar: While attention to detail is important, don't focus too much on grammar or punctuation. The important thing is.Conceptual clarityThe
• Don't underestimate the model: Don't think that models are stupid and need to be "coaxed" to work. You shouldtrust modelability and give it enough information to accomplish the task.
• Don't be afraid of complexity: As models become more capable, they can handle more complex information. Instead of trying to hide complexity, it istrust modelto deal with.
• Continuous learning and adaptation: As modeling capabilities increase, theEffective prompting methods can also become obsolete. You need to keep learning and adapting to changes in the model.
• Feedback sought. Showing your tips to others, especially those who are unfamiliar with your task, can help you spot problems you may have overlooked.
• Read the tips: Read good tips written by others and analyze how they work.

7.7 Tips for the future

• Models will be assistants: the In the future, models will help you write prompts. You will need to collaborate with the model to determine what needs to be written and what is missing.
• Increased capacity for introspection. You will need greater introspection to make yourself visible to the model.
• The point is to understand you: In the future, the focus of modeling will shift from understanding instructions to understanding your intentions.
• From Directive Recipient to Expert Advisor. The model may change from being a simple receiver of instructions to an expert advisor. You need to learn how to communicate more deeply with models and get feedback from them.

In conclusion.Tip engineering is a skill that requires practice and continuous learning. By understanding how models work, employing diverse cueing methods, and continually exploring the boundaries of the model, you can improve your cueing engineering skills and better utilize the model for a variety of tasks. Ultimately, a good cue is one that clearly, concisely, and accurately expresses your intent and allows the model to effectively accomplish the task you want.

8. Discussion on Jailbreak

What is Jailbreak?

• define: Jailbreak Prompts are prompts that attempt to bypass the security restrictions and ethical guidelines of a large language model (LLM). These prompts are usually intended to allow the model to generate content that is otherwise prohibited, such as harmful, unethical, or biased content.
• goal: The purpose of jailbreaking is usually to explore the limits of the model, to test the safety and robustness of the model, and to understand how the model responds to different inputs and wording.
• methodologies: Jailbreaking can be done in a variety of ways, including the use of large numbers of tokens, long texts, unusual wording, multilingual mixing, role-playing, and the use of models to predict the text.

How Jailbreaking Works

• Exceeding the training distribution: One possible explanation is that the jailbreak cues place the model outside of its training data distribution. For example, during the fine-tuning process, the model may not have encountered such long or complex text, and thus may behave abnormally when processing these cues.
• Utilization of forecasting mechanisms: Jailbreaks sometimes utilize the model to predict text in a way that, for example, starting a prompt with "Here's how you..." may cause the model to generate more detailed and specific responses.
• Utilizing reasoning skills: Jailbreaking may be able to get around certain restrictions by exploiting the model's reasoning capabilities, for example, by requiring the model to generate responses in other languages before translating them into the target language.
• Capitalizing on training differences: Jailbreaking may take advantage of differences in training data across languages, e.g., certain content may be allowed in one language but prohibited in another.
• social engineering: Jailbreaking sometimes smacks of social engineering, which involves not just exploiting vulnerabilities in a system, but also understanding how the system works and using that understanding to get around restrictions.
• Making Sense of Models: An effective jailbreak method requires not only trying, but also understanding how the model works, how it is trained, and using that knowledge to bypass the model's security mechanisms.

Jailbreaks and Model Training

• Purpose of model training: One of the goals of model training is to identify and eliminate jailbreak patterns so that the model can respond more securely to user input.
• Ongoing training process: Once an effective jailbreak method is discovered, the model is retrained to avoid the same vulnerability again in the future. This means that jailbreaking techniques tend to be short-term and are fixed as soon as they are discovered.
• Safety and ethics: Jailbreaking is closely related to model security and ethics. Because the ultimate goal of jailbreaking is for the model to generate content that violates security guidelines, model developers continuously iterate on the model and security mechanisms to prevent such behavior.

The Meaning of Prison Break

• Test Boundary: Jailbreaking helps us better understand the limitations of the model and improve its design by testing the boundaries of its capabilities.
• Revealing potential problems: Jailbreaking can reveal potential problems in model training, such as data bias or security vulnerabilities.
• Improved security: By studying jailbreak methods, we can develop more effective security measures that will make the model safer for practical use.

summarize

Jailbreaking is an important area of research in cue engineering that not only helps us understand how large language models work, but also helps us improve the security and reliability of our models. Jailbreaking is centered on exploring the boundaries of a model, trying to get the model to generate content it was not meant to generate, and learning and improving in the process. Jailbreaking is also closely related to the training process of the model, as the model is constantly updated and improved to eliminate potential vulnerabilities.

9. Key quotes from speakers

9.1 On the definition and nature of cue engineering:

• Zack Witten. "I think the cue project isTrying to get the model to do things, trying to maximize the potential of the model, try to work with the model to accomplish things you otherwise wouldn't be able to." He emphasized the importance of clear communication and argued that talking to modelsIt's a lot like having a conversation.The
• Zack Witten. "The engineering part comes from iterative experimentation." He points out that unlike talking to a person, talking to a model has a "reset button" that allows you to start from scratch and try different approaches independently, which makes experimentation and design possible.
• David Hershey. "I think the cue is kind of like youProgramming model approach." He noted that building a system that uses language models requires not only writing prompts, but also considering issues such as data sources, latency, and system integration.
• Zack Witten. "The articles we are writing now are just like code." He argues that written texts, such as a good essay, can now be treated like code.

9.2 On the traits of a good cue engineer:

• Amanda Askell. "I think it's a mixture of clear communication, so being able to articulate things clearly and understand tasks clearly andThink and describe concepts well." She emphasizedIterative capabilitiesas well asThink about the ways in which hints can go wrongThe
• Amanda Askell. "The Difference Between a Good Cue Engineer and a Bad Cue Engineer, is in the ability to systematically break down all the information needed for the task." She emphasized the importance of pulling from one's own understanding and moving toward modeling theCommunicating information clearlyThe importance of the
• Zack Witten. "Read the output of the model". He emphasized the importance of carefully reading the output of the model and pointed out that, even if "think step-by-step" was included in the prompt, it was important to check whether the model was actually thinking step-by-step.
• Amanda Askell. "Imistrust model, and then I just keep trying." She believes that models need to be constantly tested, especially in unfamiliar areas, to ensure their reliability.

9.3 Practices and tips on prompting:

• David Hershey. "A lot of times it's not like you write a prompt and then give it to the model and that's the end of it. In fact, it's much more than that.It's much more complicated.." He noted that prompts often need to be integrated into larger systems.
• Zack Witten. "Try not to abstract your tips(math.) genusClearly describe the task, don't try to build crazy abstractions." He argues that describing tasks clearly is usually more effective than trying to build complex abstractions.
• Amanda Askell. "The first thing I do is, I cue it up and then I'll say, 'I don't want you to follow these instructions. I just want you to tell me what is unclear about them, any ambiguities or anything you don't understand.'" She suggests asking the model to point out unclear or ambiguous areas after the initial prompt.
• Amanda Askell. "If people see a model make a mistake, they usually don'tAsk the model directly." She suggests that when a model makes a mistake, you can simply ask the model why it made the mistake and how the instructions could have been modified to avoid the error.
• David Hershey. "If you don't give it an **"exit" option**, it will keep trying to follow your instructions." He emphasized the importance of providing an "exit" option in the prompts so that the model can handle uncertainty if it encounters it.
• Amanda Askell. "Don't get overly attached to a perfect tip." She believes that overreaching for the perfect tip can lead to stagnation, and that it's important to recognize when it's time to stop optimizing.
• Zack Witten. "I usuallyTry to keep grammar and punctuation correct, because I think it's interesting." He believes that attention to detail is important, even though the model may not require perfect syntax.

9.4 On the future of roleplaying and prompts:

• Amanda Askell. "I just thoughtThere's no need to lie to them.." She argues that as the models become more powerful, there is no need to use false role-playing, and that a straightforward statement of the mission will suffice.
• Amanda Askell. "You have to put into words what you want., sometimes what I want is quite subtle." She believes that sometimes you need to invent new concepts to express your intentions and define them in concert with the model.
• Amanda Askell. "Maybe the cue will become asI explain what I want, and then the model prompts me to." She envisions a future where models can in turn prompt users to help them clarify their needs.
• Zack Witten. "I thinkWe will be using models more in the future to help us with cueing." He sees a future where models will be used to help generate cues and interact with them at high bandwidth.

9.5 On the evolution of cue engineering:

• Amanda Askell. "As time goes on, I'mIncreasingly inclined to trust itthat gives it more information and context." She argues that as models advance, they can now be trusted to handle more information and context.

9.6 Key Summary:

• Clear communication and iteration are at the heart of cue engineeringThe
• Good cueing engineers need toUnderstanding how the model worksmergeContinuously explore the boundaries of the modelThe
• In the future, models will be assistants to prompts, and can even prompt the user in turnThe
• Cue the engineering skills that will come from theTransfer of low level tasks to higher level tasks, such as task decomposition and complex reasoning.
• Introspective skills and conceptual definitionswill become even more important.

Explanation of key terms

• Prompt Engineering:A method for optimizing textual inputs (prompts) to obtain the desired output from a language model.
• Iteration:In cueing engineering, it refers to the process of continually adjusting and improving cues, each time based on feedback from the model.
• Chain of Thought:A hinting technique that requires the model to explain its reasoning process step-by-step before giving a final answer.
• Zero-Shot:Refers to the ability of a model to answer questions directly without providing any examples.
• Few-Shot:A small number of examples are provided in the prompts to guide the model output to better accomplish the task.
• Retrieval-Augmented Generation (RAG):A methodology that enables the model to access an external knowledge base to obtain relevant information when generating responses.
• Model Output:Refers to textual responses generated by the language model in response to a prompt.
• Theory of Mind:In the context of cue engineering, this refers to the ability to understand how a language model understands and processes instructions.
• RLHF (Reinforcement Learning from Human Feedback): a training technique that uses human feedback to optimize the behavior and output of a language model.
• Pretrained Model:Language models trained on large amounts of textual data and then fine-tuned for specific tasks.
• Enterprise Prompt:Tips designed for enterprise application scenarios that emphasize reliability and consistency.
• Research Prompt:Prompts designed for research purposes aim to explore modeling capabilities and obtain diverse outputs.
• Jailbreaking:An attempt to make the model generate hints of harmful or inappropriate content by bypassing security measures.
• Red Teaming:Simulate attacks to test the security and robustness of models and systems.
• Eval:A test or dataset used to measure the performance of a language model on a specific task.

Full Translation of the Podcast in Chinese

Chinese translation

Introduction (00:00-00:27)

Alex (host): Hi everyone, I'm Alex and this roundtable discussion will focus primarily on Prompt Engineering. We'll be exploring prompts from a variety of perspectives - research, consumer, and enterprise - sharing insights and discussing the nature of prompt engineering in depth.

Self-introduction by team members (00:28-02:00)

- Alex: Head of Developer Relations at Anthropic and former Anthropic Tips Engineer responsible for solution architecture and research.

- David Hershey: Primarily responsible for working with clients to help them fine-tune their models and solve common problems when adopting language models, such as prompt engineering and building systems based on language models.

- Amanda Askell: One of Anthropic's fine-tuning team leaders, dedicated to making the Claude More honest and friendly.

- Zack Witten: Anthropic Prompting Engineer who has worked with clients and is currently working to elevate prompting engineering throughout the community, such as developing prompt generators and various educational materials.

What is a cue project? (02:01-06:29)

Alex: What is a cue project? Why is it called a "project"? What exactly is a "hint"?

Zack: Cue engineering is designed to guide models through tasks, to fully utilize their potential, and to accomplish work that could not otherwise be done by collaborating with them. At its core is clear communication. Talking to a model is similar to talking to a person in many ways, and requires an understanding of the model's "psychology".

Alex: Why does it have "engineering" in the name?

Zack: "Engineering" is embodied in the process of trial and error. Unlike people, models can "start over", meaning you can try different approaches from scratch and avoid interfering with each other. This ability to experiment and design gives cue engineering its "engineering" properties.

Alex: So cue engineering is the process of writing cues, interacting with the model, iteratively modifying it, and being able to fall back to the initial state each time, a process that is itself "engineering".

Zack: Another aspect is the integration of prompts into the overall system.

David: Hints can be seen as a way to write a model, but more important is clarity. Thinking of it as programming requires consideration of data sources, accessible data, latency tradeoffs, and the amount of data provided. Building models requires systematic thinking, and this is what distinguishes cue engineering from software engineers or product managers; it is self-contained.

Alex: Are hints natural language code? Is it a higher level of abstraction or a separate concept?

David: Over-abstracting hints can complicate the problem; usually only a clear description of the task is needed. However, hints do compile instructions into results, so important concepts in programming such as precision, version control, and experiment tracking apply to hints as well.

Zack: Now, it makes sense that we treat well-written articles as if they were code.

What qualities should a good prompting engineer have? (06:30-12:43)

Alex: What makes a good cueing engineer?Amanda, what do you look for when hiring a research cueing engineer?

Amanda: Good cueing engineers need to be able to communicate clearly, iterate, and anticipate situations where cues might go wrong. Clear communication means being able to articulate, understand tasks, and describe concepts accurately. Excellent writing skills are not entirely correlated with excellent cue engineering skills. Cue engineering does not happen overnight; it requires constant iteration to analyze the model for areas of misunderstanding and make corrections. Good cue engineers think about specific situations where the model might be wrong, such as a dataset that doesn't have a name that starts with "G" or an empty input string, and add explanations for those situations.

David: Engineers often consider ideal situations in which a user might type, but the reality may be that the user does not use capital letters, misspells, or enters meaningless words. Being able to anticipate actual user behavior is another important ability of cueing engineers.

Zack: Reading the output of the model is critical. Similar to "looking at the data" in machine learning, cue engineering requires careful reading of the model's output. For example, even if a cue asks the model to "think step-by-step," it needs to be checked that the model actually does so, as the model may understand the instructions in a more abstract or generalized way.

Amanda: Writing a mission statement is very difficult and requires clearly communicating information that Claude does not know. Many people will write down the information they know straight away, but without systematically sorting out the complete set of information needed to understand the task.

David: Many people write prompts that are based on their a priori understanding of the task, rendering them incomprehensible to others. Good cueing engineers are able to step outside of their own knowledge framework and communicate the task in its entirety to the model.

Alex: Often times, based on prompts written by others, I am unable to complete the task while the model is expected to do a better job than me.

David: Current models are not yet able to ask targeted questions in the same way that humans do, so prompting engineers need to think for themselves about what questions the other person might ask and answer those questions in the prompt.

Amanda: I would have the model point out any unclear or ambiguous parts of the prompt and have the model explain what went wrong and suggest changes.

How can I tell if a model can detect its own errors? (12:43-14:12)

Alex: Can a model really discover its own mistakes by asking "why did it make a mistake"? Are the explanations it provides real or are they "illusions" about the model's own capabilities?

Amanda: If you explain to the model what it is doing wrong, it can sometimes recognize the problem. But it depends on the specific task and the success rate is uncertain, but I always try.

Zack: Interacting with the model can help you understand the situation, and you'll miss out on this information if you don't try.

How can I tell if a prompt is credible? (14:13-17:52)

Alex: You interact with Claude a lot on your Slack channel and use it for various research scenarios. How did you build trust in the model?

Amanda: I don't trust the model completely, but I "hammer" it constantly. I think "Can I trust you to do this?". I would think "Can I trust you to do this task?". Models are sometimes unreliable on seemingly simple tasks, often in areas outside the distribution of the model's training data. This is decreasing as models become more capable. I don't trust models by default, but I think that in machine learning, one usually wants to look at a lot of data to eliminate noise. And in cue engineering, a small number of carefully constructed cues is more valuable than a large number of randomly constructed cues. If I look at the outputs of 100 models and the results are consistent, and I know that the outputs cover a wide range of edge cases and anomalous inputs, then I'm going to trust the model more.

David: Signals in machine learning are usually numbers, such as prediction accuracy. And the output of a model is usually a large amount of text, from which we can learn how the model thinks. It's not just whether the model did the task correctly, but how it arrived at the result and what steps it went through.

Amanda: Well-written hints can boost experimental success from 1% or even 0.1% to the top 1% or even the top 0.1%. If your experiments need to be in the top 1% of model performance rankings in order to be successful, it is critical that you spend time on the hints.

David: In product deployment, a good tip can make an otherwise unreleasable product usable.

Amanda: But there is also the trap of "better tips are always on the way".

How can I tell if a task can be solved with a prompt? (17:53-21:12)

Alex: How can I tell if a task is likely to be solved by a prompt?

Amanda: I usually check if the model "understands" the task. If it is clear that the model is unable to accomplish a task, I don't spend much time on it.

David: You can guide the model to state its thought process and from there determine if it has understood the task correctly. If the model's thought process is completely different each time and is far from the right direction, I usually give up.

Amanda: Now, this is rare.

David: I recently tried to get Claude to play Pokemon, which was the first time I've ever encountered this. I spent a weekend writing hints to try and get Claude to understand the Game Boy screen, and while I made some progress, it wasn't enough. So I decided to give up for now and wait for the next model.

Tips on images (21:13-24:27)

Zack: One of the things I liked about the hints you used in the Pokemon game was that you explained to the model that it was in a Pokemon game, and how the game elements were represented.

David: I ended up superimposing a grid over the image and describing each grid section, then reconstructing it as an ASCII drawing with as much detail as possible. This has a lot of similarities to the cue project, but I've never done this with an image before. I found that many of my intuitions about text did not apply to images. For example, multi-sample cues don't work as well on images as they do on text.

Alex: We previously found it difficult to improve Claude's perception on images when exploring multimodal cues.

David: I was eventually able to get Claude to recognize walls and characters most of the time, but it couldn't recognize NPCs, which is crucial to playing the game well.

Discussion on role-playing prompts (24:28-32:26)

Alex: Is the cueing technique of telling the model that it plays a certain role or identity effective?

Amanda: As models become more capable and better understood, I don't see the need to lie about them. I don't like lying, and I don't think building an assessment dataset for a machine learning system is the same as building a quiz for children. The models know what language model assessment is, so I'll prompt them directly for the actual task. I'd tell the model "I want you to construct questions that are very similar to language model assessments" rather than pretending to complete an unrelated task.

Zack: I've found that using metaphors can help the model understand the task. For example, when judging the quality of a chart, I ask the model "If this were a high school assignment, how much would you grade this chart?" . This doesn't mean "you're a high school teacher," but rather provides an analogy that allows the model to understand the way I expect the analysis to be done.

David: People often use role-playing as a shortcut to accomplish similar tasks, but they don't realize how much product detail is lost. As models become more capable, it is more important to more accurately describe the specific context in which the model will be used. For example, instead of saying, "You are a helpful assistant," tell the model, "You are in this product, you represent this company, and you are the support chat window for this product. My advice is to describe the specific context in which the model will be used in as much detail as possible, as people often get sidetracked from the actual task by role-playing.

Amanda: Personally, I've never used roleplaying as a prompting technique, even on less capable models.

David: This may be related to the differences between the pre-trained model and the RLHF model.

Amanda: I'd visualize the task as being done by a temp, and I'd tell him "We want you to detect good charts, and by good charts I mean ......", but I wouldn't say to him "You're a high school student! ".

Suggestions for concise presentation (32:27-36:45)

David: When clients say their prompts don't work, I ask them to describe the task and then have them record what they just said and transcribe it into text, which is usually better than the prompts they wrote.

Zack: Someone asked us to help optimize the tips, so I just copied what they described and the tips worked.

David: People haven't fully understood what prompts really mean. Many people use text boxes as Google search boxes to enter keywords. In enterprise applications, people try to take shortcuts in prompts, thinking that a particular line of text is important. People put a lot of effort into finding the perfect, insightful sentence, but it's hard to do.

Amanda: People often forget to leave room for the model in their prompts. For example, the model will do its best to follow your instructions when it encounters an edge case, but if you don't tell it what to do, it may give the wrong answer. You can tell the model "If something weird happens and you're not sure what to do, output 'not sure' in the label". This will help you identify situations that the model is not handling well and improve the quality of the data.

Amanda: I would show the prompts to others as if I were doing the assessment myself.

David: Karpathy also makes its own ImageNet test sets.

How to get valid information from model responses (36:46-40:46)

Alex: How do you get valid information from a model's responses? It's not just a number, you can learn about the model's thought process from it. Does this apply to thought chains?

David: I think the personification analogy, with its overemphasis on "reasoning", is harmful. The important thing is that chains of thought do work and can improve model performance. Structured reasoning can further enhance the effect.

Amanda: If you remove the reasoning process by which the model arrives at the right answer and replace it with reasoning that seems reasonable but leads to the wrong answer, see if the model comes to the wrong conclusion.

Zack: Having the model write a story before completing the task doesn't work as well as a chain of thought.

Alex: This suggests that the reasoning process does have an impact on the outcome.

Amanda: I've seen cases where the steps of reasoning don't agree but end up with the correct answer.

On the need for grammar and punctuation in prompts (40:47-45:19)

Alex: Does the prompt require attention to grammar and punctuation?

Zack: I'll pay attention to these details because it's fun, but it's not necessary. What is important is that you should have that attention to detail.

Amanda: I often make spelling mistakes in my prompts, but I'm more concerned with the clear expression of concepts.

David: This is related to the pre-training model and the RLHF model. The conditional probability of spelling errors is higher in the pre-trained model. Applying the intuition of the pre-trained model to the model in the production environment does not always work.

Alex: Talking to the model can be seen as a form of mimicry to some extent.

David: The model adjusts its behavior based on your inputs.

The Difference Between Business Tips, Research Tips, and General Chat (45:20-50:53)

Alex: What's the difference between business tips, research tips and general chat?

Zack: Research-based prompts are more focused on variety and exploring the possibilities of the model, and therefore have fewer or no examples to avoid over-reliance on examples for the model. In contrast, enterprise-level prompts are more focused on reliability and format consistency, and therefore use a large number of examples.

Amanda: The examples I use are usually different from the data the model will be working with, and are intended to illustrate concepts rather than make the model memorize them. For cognitive tasks, I want the model to actually understand the correct answer in each sample.

David: On Claude.ai, I only need to get the model to complete the task correctly once. But in an enterprise application, the prompts need to be able to respond to a variety of situations and input data.

Suggestions for improving cue engineering skills (50:54-53:57)

Alex: Suggestions for improving tip engineering skills?

Zack: Read excellent tips and model outputs, analyze the principles and experiment with them, and talk to the models more often.

Amanda: Show your tips to others, especially those who don't know your work. Keep practicing and look at your tips from a "beginner's" point of view.

David: Try to get the model to do something you don't think it can do.

About jailbreak (53:58-56:54)

Alex: What happens inside the model when people write jailbreak hints?

Amanda: One possibility is that the jailbreak cue biases the model away from the distribution of the training data.

Zack: Jailbreaking sometimes seems like a combination of hacking and social engineering.

Evolution of cue engineering (56:55-64:33)

Alex: How has the cue project changed over the past three years?

Zack: We will incorporate effective cue engineering techniques into model training, so the best techniques are usually short-lived.

David: I've gradually learned to respect the models' ability to give them more information and context.

Amanda: I would give the paper directly to the model and let it learn the cueing technique on its own.

David: People often underestimate the power of models and try to simplify the problem to "Claude's level".

Amanda: I will try to get into the "mindspace" of the model, which will affect the way I write the prompts.

Zack: It's easier for me to get into the mindspace of a pre-trained model.

Amanda: Reading content on the Internet may be more helpful in understanding the model than reading a book.

Tips for the Future of Engineering (64:34-end)

Alex: What is the future of cue engineering? Will we all become cue engineers?

David: Specifying the goals of the model is always necessary, and it is important to express them clearly. Tools and methods will continue to evolve, and models can help us write better prompts.

Zack: We will make more use of models to assist in prompting the project, e.g. to generate examples.

Amanda: I currently write mostly meta-prompts that allow the model to generate the output I want. In the future, the model may act like a designer, interacting with us and guiding us to say what we really want.

David: I'll have Claude "interview" me to extract the information.

Amanda: Right now, we need to communicate the concepts in our minds to the model, and in the future, the model may actively guide us to say those concepts. Philosophical training helps me to clearly express complex concepts.

Alex: Extracting information from users will become even more important.

Zack: Cue engineering is like teaching, you need to "empathize" with your students. In the future, we need to "introspect" and let the model understand us.

Amanda: I often define new concepts to clearly express my ideas.

Alex: Amanda sums it up perfectly: externalizing your ideas to an educated layman.

Summary:

This roundtable discussion is centered around cueing engineering, covering a variety of aspects such as its definition, the qualities of a good cueing engineer, how it interacts with models, enterprise-level applications, research applications, and future directions. Core points include:

- At the heart of cue engineering is clear communication and an understanding of modeling capabilities.

- Good cueing engineers need to be able to articulate clearly, iterate, anticipate errors, and think systematically.

- As the model's capabilities increase, cue engineering will focus more on how to extract information from the user, rather than issuing commands to the model in a one-way direction.

- The future of cue engineering may resemble interactions between designers and clients, with models taking a more proactive role in guiding users to express their needs.

- Philosophy training helps to improve cue engineering skills because philosophy emphasizes the clear and accurate expression of complex concepts.

The layman's guide to getting started with AI

Help you learn how to utilize AI tools at a low cost and from a zero base.AI, like office software, is an essential skill for everyone. Mastering AI will give you an edge in your job search and half the effort in your future work and studies.

View Details>