🤖 How to build a fully automatic LP robot with AI "mouth massage" at 3 costs Yesterday, we mentioned an "Automatic LP Strategy Idea". Although the original document has some loopholes, it does not hinder us from providing a good idea, which is worth exploring in depth. So, he drew his bow and arrows, and the code's mouth rolled up. Due to the fact that this is not the kind of "one-time" code, with many functions and relatively new content, it is not possible to have an AI help you generate a complete and usable system in one or two sentences. So, my process is as follows: one ️⃣ O3 production PRD The first step is not to directly tell AI what software you need and what functions to implement. And first, communicate the requirements with a highly intelligent AI, and on this basis, let the AI help you generate a software PRD (Product Requisition Document), which is the product requirements document. Attention, first communicate the requirements and then have AI generate a detailed PRD document for you. I chose OpenAI O3 to help me complete it. two ️⃣ Prepare materials The so-called clever woman cannot cook without rice. For relatively new fields, the knowledge base of AI may not have the latest content, so you need to do some preparation work. For example, I found the official documentation of Meteora and downloaded the introduction and SDK documents about DAMM V2 as Markdown text, keeping it as a backup (for AI, it is easier to process and understand markdown, but if it is in PDF format, it is more difficult and prone to errors) three ️⃣ Claude Code is playing I chose Claude Code, an AI programming tool that runs on the command line (and specifically re opened Claude Pro membership for this purpose). In the script directory you have chosen, place the prepared document in docs/and inform Claude, then reference the PRD document you prepared in the first step. Let Claude Code start building the framework of the entire software based on the documentation information in the docs (ensuring knowledge of the specific usage of the SDK) and the PRD document. Last night, I used Claude Code through API access, and the entire architecture was built at a cost of only 3. I wrote about 2600 lines of code. four ️⃣ Cursor seam repair The code generated for the first time, 99% of the time, cannot run successfully, there must be bugs waiting for your AI to fix. This step could have been completed directly within Claude Code, but some error locations are difficult to describe using the command line. So, I opened the cursor, selected Claude 4 Sonnet, started pointing and fixing bugs wherever I went. After a few rounds, the program can now run happily (completed Jupiter Swap integration and successfully tested). The DAMM V2 access function still needs further testing. five ️⃣ GitHub Integration For a relatively large project, it is essential to integrate the code into GitHub (or a similar code hosting platform) and enable a private repository (invisible to others). The biggest advantage of this is that every time a feature is added or fixed, it can be committed in a timely manner. In this way, not only is every modification to the code clear, but most importantly, it can be rolled back to any previous available version in a timely manner when problems are discovered. Moreover, with Git information, IDE tools such as cursor and VS Code can quickly compare the latest modifications. With the support of the strongest programming AI on the surface, such as O3, Claude 4, and Gemini 2.5 Pro, with the above five steps, it is basically possible to use a completely usable tool. Even a tool that automatically helps you participate in @ MeteoraAG LP is no problem. Below are a few screenshots for your reference