Skip to main content

Turning AIs Into Cavemen Is Actually A Good Idea

 

 

Token limits have been the number 1 problem of vibecoding. Aside from high usage, vague prompting and inefficient context management are two key causes that contribute to this problem. Tokens are counted from two things, the user’s input and agent’s output. You can have the best prompt in the world, but if your AI puke a Shakespearean essay for each response, your token will reach its limit faster. Similarly, if your prompt is too long, your token usage will also skyrocket.


Prompting

Prompting seems easy enough, but the reality is, if you want quality output, the input must be of similar quality. Here are some tips you can try to achieve a better, token-efficient prompt.

Stick to English language only

Certain languages may use higher tokens. Switching between languages might not be a good idea. Sure you can control your input prompt, but the output may use more tokens.

Extremely short input

Do not use ‘please’ or ‘thank you’. Get rid of filler words. Straight to the point. Talk like a caveman. ‘I want to make a fire by heating up wood, thanks.’ is bad, ‘Make fire with wood’ is good. 

Extremely short output

You caveman. Is agent caveman too? Install the caveman extension to make sure the output is always just as short as the input. Installing the extension is easy, and it's properly documented there, so go check it out.

Do not refactor outputs

Why ask your agent to refactor a code they just gave you? Instead, make sure your prompt is clear enough so the output code doesn’t need to be refactored afterwards.


If you want to check how many tokens your prompts are using, verify it on this tokenizer website 


Context Management

Set reference points

Mention files to refer (eg: @file_path/file) and mention functions to refer. Do not directly paste code into the prompt input, mentioning is better. Code outputs that are based on existing code always use less token than a completely new code output. 

Long context is expensive

Start a new session for a new task. Clear conversation once you’re done. Regularly review context usage with /context command. 

Use context management engines

This one is a bit ‘open’. There is no one-solution-fix-all here. Some use a custom made skill, some use extensions, some use softwares. It’s completely up to you. The key is to make sure your agent has just enough context to ensure good token efficiency while still having a good response. One tool that I use is Vexp.


TLDR

Why use many words when few do tricks? Embrace caveman conversation to use fewer tokens. Hope this post helps your vibecoding sessions.

Comments

Popular posts from this blog

🗑️ Clear storage Mac OS

  🗑️ Clear storage Mac OS 1: Clear system cache: Go to Finder > Go > Go to Folder, then type in "~/Library/Caches" and hit enter. Select all the folders inside the Caches folder and delete them. 2: Clear system logs: Go to Finder > Go > Go to Folder, then type in "/var/log" and hit enter. Select all the files inside the Log folder and delete them. 3: Remove unused language files: Go to Finder > Go > Go to Folder, then type in "/Library/Languages" and hit enter. Delete all the language folders you don't need. 4: Uninstall unused apps: Go to the Applications folder and delete the apps you don't use. 5: Clean up system files: Use a system cleaning tool like CleanMyMac X to scan and remove unnecessary system files. 6: If you have npm installed, clear the caches once in a while with ‘sudo npm cache clean --force’ 7: If you have ionic projects, open the ‘.angular’ folder and delete the ‘cache’ folder inside it.

Git Revert is Easier Than You Might Think

    Its surprisingly easy and not as complicated as some people might say. You just need to know which one is the 'bad' commit, then revert to one commit before the 'bad' one. In this example, person 'M' accidentally pulled from the dev branch instead of main. So to revert this, just follow these steps: Find the 'bad' commit In this case it's git message have something like 'pulled dev into branch...'. This is the bad commit, do not copy the git hash of this commit. Instead, get the last 'good' commit. Copy the commit hash of the one before the 'bad' one. Paste it somewhere safe, we will get back to this later. Do not revert right now.   Backup your commits after the 'bad' one After pulling from dev, 'M' pushed a few more commits afterwards. If we've already reverted, finding the committed changes would be a bit harder. So now is the best chance to copy all the commit hashes for cherry-picking later.    Rev...