Increase your llm exit and design sharp promotion: Real tactics from the Ai Legeer's box
but Gossip provisions Operating and reliable exit is not. Since language models grow strongly and flexion, to get the highest results depending on many How do you ask The model is model itself. This is where the engineer is immediately, not as ideasical exercise, but as Daily Information Active In production locations, and thousands of phone calls daily.
In this article, I share Five operating strategies of effective engineering I use almost every day to create a stable and reliable, higher AI. It is not just tips I have learned about but the means I tested, analyzed, and depends on all real use of the world in my work.
Some may seem tricky, and others are surprised, but they all make a real difference my artifacts to find the results I expect from the LLMS. Let's get in.
Tip 1 – Ask the llm to rewrite
This first approach can feel contemplatedBut it is one time I use. Instead of trying to make the passing canoes from the beginning, I usually start with a difficult frame of what I want, then ask llm that Analyze the ideal Prompt for whilebased on additional context I offer. This Building Strategy allows immediate production of Direct and effective motivation.
The perfect procedure is usually constructed in three steps:
- Start with a normal structure to explain the rules to follow
- Itrative test / recyclation of later to match the desired result
- General integration of cases on the edge or specific requirements
When the llm lifts quickly, I run it by a few Regular examples. If the results are turned off, I don't just finish my hand. Instead, I ask a llm to do that, a specific question General repairsAs the llms often add items in a very specific way. When I get the answer you want 90% +, I usually drive it in Batch of the installation data Analyzes of refugees need attention. Then I brought this problem to the problem when it included input and Ouput, for Tatatative tendency.
A good tip that usually helps too much Requiring the llm to ask questions Before lifting quick conversion to move the need.
So, why is this effective?
a. Ed quickly.
Especially of complex tasks, the llm helps make up the problem space in the way you are logical including operate. It helps me Explain my thinking. I avoid getting on the floor in syntax and always focuses on solving the problem itself.
b. Reduces conflict.
Because the llm translates this activity by their names in the «itself», more likely to get mysterious or argue. And when it does, it often asks for clarification before raising clean, irrefutable shape. After all, I have better to call a message than The one intended to translate?
Think about it as interacting with a person: a large part of the Miscumunication appear different tasks. Llm gets sometimes invisible or arguing so that I thought it was completely seen … and finally, it's the work, so definite Important translation, not mine.
c. Even better.
Sometimes I strive to get clear, mysterious formulation of work. The llm is wonderful in this case. They see the pattern and produce quickly normal and firm in what I could.
Tip 2 – Use Self-Examination
The idea is simple, but also, he has great power. The goal is to force a llm to check the quality of its response Before you take it out. Straight, I ask measure its response with a pre-defined average, from 1 to 10. If the points are under a certain limit (usually I put it at 9), please dispute or improve Answer, depending on work. Sometimes I can add an idea of ”if you can do better” to avoid endless loop.
In fact, I find it interesting that the llm usually behaves in the same way people: usually The easiest answer is the best one. After all, the LLMs are trained for people generated by people so that they are not designed to repeat the answers patterns. So, to give it Clearest of normal It is very helpful to improve the final result of exit.
The same method can be used for a Average Starting Last focused on receiving the law. The idea is to ask the llm to review its response and ensure that it follows a particular law or all rules before submitting an answer. This can help improve the quality of response, especially when one law often flies sometimes. However, with my experience, this option works well than requesting the quality score. When this is required, maybe it means that your immediate or ai movement requires improvement.
TIP 3 – Use Reply structure and an intended example to integrate format and content
Using examples is a well-known and powerful way to improve results … As long as you don't use too much. Well-chosen example usually helps more than many educational lines.
Answerance, on the other hand, helps to explain how the result should be viewed, especially technical or repetitive activities. It avoids wonders and keeps the results move.
For example and fill the building by showing how we can fulfill your consideration. This «Building + For example» Combo usually works well.
However, examples are often Text-heavyand to use the most than all can Transfer the most important rules or lead to them are consistently followed. They also increased the number of tokens, which can create side effects.
Therefore, use examples wisely: one or two well-chosen examples cover your important rules or edge is usually sufficient. Adding more may not be worth it. Can help and add a A brief description After example, why to justify the application, especially if that is really invisible. I personally rarely use negative examples.
I usually give examples of domestic or two and the general composition of the expected output. Most of the time I choose XML tags
Providing an example is especially useful when the building is limited. It makes things very clear.
## Here is an example
Expected Output :
-
My sub sub item 1 text
My sub sub item 2 text
My sub item 2 text
My sub item 3 text
-
My sub item 1 text
My sub sub item 1 text
Explanation :
Text of the explanationTip 4 – Break Complex Tasks to Simple Steps
This may seem obviously, but it is important to keep the_the_the_the_the_the Quality. The idea is that Different a great work on a few small stairs, well-defined.
As a person's brain struggling when there should be a multitask, llms often produce low-quality responses where the work is very wide or involves very different purposes. For example, if I ask about 125 + 47, and 256 – 24, and at end 78 + 25, one after another, this should be good (hope :). But if I ask you to give me three answers by looking at one, the work becomes more complex. I like to think of that llms treat the same way.
So instead of asking the model to do everything like reading an article, translating it, and trashing the process into two simple or three-handed steps.
The lower mate of this method is to add specific difficulty into your code, especially when it exceeds information from one step to the following. But modern structures such as LangchainWhat I love and use whenever I have to deal with this situation, do this kind of work control in the simplest order to use.
Tip 5 – Ask the llm to describe
Sometimes, it's hard to understand why Llm gave an unexpected answer. You can start guessing, but the easiest and most honest approach may just be Asking the model to explain its thinking.
Some may say that the llm predictor does not allow the llm to describe its thinking because it simply makes – The reason but my experience shows that:
1- Most of the time, it will be successful sensible description that expressed its response
2- Making quick conversion according to this explanation usually corrects the wrong reaction of the llm.
Of course, this is not evidence that the llM actually shows, and It's not my job to prove thisBut I can say that this solution works in getting very well in good working.
This method is especially useful during the development, pre-production, or even the first weeks after life. In many cases, it is difficult to expect all the crimes that rely on one or several llm calls. Being able to understand why The model that produces a specific response helps you design a very direct adjustment that may, solve the problem without causing unwanted side effects elsewhere.
Store
Working with the LLMS is like working with a citizen trained, faster and capable, but it is usually dirty and going to all the way you expect. Finding the best in Internal requires clear commands and administrative experience. Same is the same with the llms that Wise rescue and experiences make all the difference.
Five Techniques that I have shared above non-“magical tricks” but Practical ways I use every day to go In addition to normal results Received with a regular lift process and get the high quality I need. They always helped me to answer the right results to be good. Even if it is a promotion of model design, breaking jobs are controlled parts, or simply asking the llm why The answer is, These strategies turn essential tools in my daily work to use the best Ai boats.
Quick improvement is not just by writing clear and orderly commands. It is about understanding How Translation Model They also design your way. Fast development is on the way of art, one of the nuean, a fine, and personal style, where the two composers are writing the same lines. After something else, one thing is always true with llms: Better talk to them, work for you better.
