Guidelines

These guidelines explain how to get the most out of AI Token Cleaner and how to use it responsibly. They sit alongside our Terms and Usage Policy — where those set the rules, this page is the practical “how to use it well.”

What AI Token Cleaner does

AI Token Cleaner strips the bloat from your LLM prompts — removing redundancy, hedging, and filler so the same instruction costs fewer tokens, without losing your intent. You paste a prompt, pick a mode and a model, and get a leaner version back.

What you can clean

AI Token Cleaner works on text prompts you intend to send to an LLM — instructions, questions, system prompts, agent instructions — up to your plan’s character limit (2,000 on Guest and Free, 10,000 on Pro). It cleans and compresses your prompt; it does not answer it, run it, or generate new content for you.

Please don’t submit content that is illegal, harmful, or that you have no right to use, or that contains secrets or other people’s personal data. The full rules are in our Usage Policy.

The three modes

• PlainClean — trims redundancy and hedging while keeping natural language. Best for everyday prompts you’ll still read yourself.
• StructuredClean — strips redundancy and converts your prompt to JSON or XML. Best for prompts you pass programmatically through pipelines, agents, or APIs.
• CaveMan — drops articles, filler, and pleasantries, leaving raw action-first intent on a single line. Best when you want the absolute minimum tokens and the model only needs the core instruction.

Choosing your models

Two selectors in the playground do different jobs:

• Cleaner model — the AI that actually performs the clean. Pick any supported model; premium models require Pro.
• Target Model — where you plan to use the cleaned prompt. This only adjusts the token-savings figures shown (different models count tokens differently) — it does not change which AI cleans your prompt.

Getting good results

• Give the cleaner a complete prompt, not a half-sentence — it trims what’s there, it doesn’t invent missing intent.
• Pick the Target Model that matches where you’ll actually use the prompt, so the token savings shown reflect your real model.
• Compare the original and cleaned versions (use “Highlight Removed”) before you rely on the output.
• For critical prompts, prefer PlainClean over CaveMan — aggressive trimming can drop nuance.

Responsible use

• Don’t paste secrets or sensitive personal data. Prompts are sent to third-party language models to be cleaned — treat them like any external API. Avoid passwords, API keys, financial data, or other people’s personal information.
• You’re responsible for your prompts. You should have the right to use any content you submit.
• Verify the output. AI models can make mistakes — check a cleaned prompt before using it in production.

Acceptable conduct

• Use the service for lawful purposes only.
• Don’t attempt to abuse, overload, reverse-engineer, or bypass the limits of the service.
• Don’t use it to process content that is illegal, harmful, or that you have no right to use.
• The formal, enforceable rules live in our Usage Policy — those govern if anything here is unclear.

Getting help

Questions, problems, or feedback? Reach our support contact (the same one listed in our Terms and Privacy Policy). We’re happy to help.