Claude Code: Optimizing Prompts for Maximum Efficiency

LLM Applications

Maximizing Efficiency with Claude Code

In the world of AI-assisted programming, the way you formulate your requests, or "prompts," to coding agents like Claude Code can make a significant difference. Although large language models (LLMs) have become extremely powerful, the accuracy and efficiency of their responses still largely depend on the quality of the prompts you provide.

Why Optimize How You Formulate Your Prompts

First of all, you might think that the way you write your prompts no longer matters, as LLMs have become so powerful that they can understand you anyway. To some extent, I agree with this idea. However, there are still good and bad ways to formulate prompts.

A good prompt would be explicit about what you want, include how to test the implementation, and clarify any ambiguities. In contrast, a bad prompt might make assumptions that do not align with the existing codebase.

Moreover, formulating prompts is not just a one-time action you perform for a specific thread. It is also the science of working iteratively with the model to establish a plan for implementing a feature. Thus, a good prompt might also include elements like asking the model to uncover ambiguities and discuss them with you or indicating what assumptions to make.

How I Formulate My Prompts for Claude Code

I will now discuss how I specifically formulate my prompts for Claude Code. I will explain some techniques I use to get the most out of Claude Code, although these techniques are quite general and also apply to other coding agents, like Codex.

The first topic I would like to address is that I use a transcription tool to convey all my thoughts to the coding agents. Currently, I use a transcription tool called FluidVoice, which is a model you can run locally for free and is incredibly efficient, both fast and accurate for transcribing text in English.

Transcription tools are obviously very valuable as they are faster than manual typing. Of course, speed depends on your typing ability, but it is reasonable to assume that you can speak, on average, two to three times faster than you can type on your computer. Additionally, it saves your hands from fatigue and allows you to focus on the content rather than the typing.

Another advantage I have noticed is that I am more inclined to include all the context my coding agent needs to perform a task. When I had to type everything by hand, there were times when I didn't want to write out all the context simply because it took too long. This is, of course, very detrimental because if you do not provide Claude Code with the necessary context to solve a task, it will be very difficult for it to do so. However, now that I transcribe instead of type, it is much easier for me to include all the necessary context, and it requires very little effort.

Overall, transcription tools are highly beneficial as they save you a lot of time and make it easier to include all the context you need to provide to the coding agent for it to perform the task effectively.

Indicating to the Model How to Test

Another very important point is to indicate to the model how to test. Suppose you ask the model to implement feature A. That's fine, but you also need to tell it how to test that feature. For example, if you are implementing a chat feature in your application, you need to instruct the model to open the chat in Chrome, type a query, and ensure that the AI responds correctly and that the token flows work.

Basically, you need to tell the model how it can know that it has correctly implemented the feature. This will save you a lot of time, as you won't have to test the feature yourself as much, and the model will present you with a more operational feature immediately.

Having a Discussion with the Model

A common misconception regarding prompt formulation for coding agents is that you simply provide a single prompt and let the model work until it is done. If you have the perfect prompt setup, this is possible, as agents can work for many hours and accomplish a lot of tasks. However, this does not take into account how to build the perfect prompt.

For a human, it is practically impossible to construct a perfect prompt because you, as a human, cannot have the complete context of the entire codebase. There is simply too much code, too many things to consider, and a context you might forget, which is not possible to retain in your short-term memory. However, coding agents are capable of storing all of this in their memory, making them a very good discussion partner.

Basically, what you should do every time you start with a coding agent, whether to implement a new feature, resolve a bug, or something else, is to have a discussion with your coding agent. You should ask your coding agent to research the topic, check the logs, look online for anything relevant, etc. And ensure that the agent asks you all the questions about ambiguous elements.

When coding agents write code for you, they make a ton of assumptions. Many more assumptions than you are even aware of. The reason is that there are too many decisions to make in coding for the coding agent to ask you every time. When humans code as well, they have to make a number of assumptions, sometimes conscious, sometimes unconscious. However, during planning, you want the model to share at least the significant assumptions it is making.

An example of such an assumption could be whether to implement the infrastructure directly via a CLI, for example, for AWS, or to implement it as a CDK, thus as code. This is not a perfect example, as in almost all cases, the right answer is that you should implement it as code and not directly via the CLI. But it is an example of a decision you, as a human, want to be aware of so that the coding agent does nothing you wouldn't want it to do. You essentially have this discussion with your coding agent to align the implementation in your mind with what the coding agent is actually going to implement.

HTML Reports

One last thing I want to address in this article is HTML reports. When coding agents work, there is a lot of information. The agent will write a lot of output in your terminal, most of which does not concern you, such as tool calls or just reflection tokens, etc. However, sometimes you need to examine what the coding agent is doing. This could, for example, be the plan it established during the last task where you want to have a discussion with the model about how to implement a feature. In these cases, you should almost always ask the model to create HTML reports to visualize what you are discussing, which greatly facilitates providing your feedback to the model.

Let's be very specific here. Suppose I am in the process of implementing feature A, but there are several decisions the model wants to discuss with me before starting the implementation. In this case, what I do is ask the model to generate an HTML report with each decision listed, including context such as:

• What are the different options for this decision
• What the decision is about
• Pros and cons of the decision

This makes it incredibly simple for me to understand what the different options are for each decision and to have an overview of the decisions to be made. And it is much easier than reading and understanding that through the terminal, for example. Another advantage of HTML reports is that you can create visuals, such as diagrams, flowcharts, SVGs, etc.

I also create HTML reports for testing. So, let's say the model has implemented a feature I requested or fixed a bug. I ask it to create an HTML report containing each part of the feature or bug, with details on how it was tested and the results of the tests. This allows me to clearly see what has been done and how it works, which is essential to ensure everything functions as expected.