Claude Was Eating 84.6K Tokens at the Start of Every Chat Until I Made This One Change

Claude Was Eating 84.6K Tokens at the Start of Every Chat Until I Made This One Change

Are your Claude chats burning through tokens before you even ask anything? One user discovered that a brand-new chat had already consumed 84.6K tokens after a single “Hi” — and a quick cleanup of what loads at the start of every chat cut that number down by around 16K tokens. Here is what was eating the context, and what actually helped.

Claude app (Credits: How-To Geek)

Why Claude uses 84.6K tokens before you ask anything

The discovery happened by accident. After switching from Claude’s Cowork mode to Code mode in the desktop app — where token usage is visible in the right-hand panel — the author started a fresh session, sent a simple greeting, and was stunned to see 84.6K tokens already consumed out of a 1M-token context window.

The breakdown told the story. About 30K tokens came from Claude’s system prompt and built-in tools, which are baked into the app and cannot be removed. Another 23K tokens sat under Messages — far more than a two-line greeting could explain — which the author believes includes memories and context from past conversations loaded at the start of every chat. The remaining 30K or so came from installed Skills, Plugins and MCP tools.

The one change that cut Claude token usage by 16K

The fix was a two-part cleanup of everything that loads before you type. First, the author deleted old conversations, which dropped the Messages chunk from around 23K to around 14K tokens. Disabling the memory feature could have cut it further, but the author chose to keep it because it is genuinely useful.

Next came the Skills and plugins. Going into Settings and disabling almost every Connector, Plugin and Skill saved around another 8K tokens. Interestingly, going scorched-earth and deleting them entirely made no extra difference — disabling them was enough to remove them from the context. All told, the cleanup brought the starting cost down by roughly 16K tokens.

The honest takeaway: you cannot bring the number to zero. The roughly 30K system prompt is fixed, and some overhead is the price of a capable assistant. But if you have been installing Skills and plugins for months without pruning them, or hoarding years of old chats, a five-minute audit is the fastest way to stop paying tokens for tools and memories you never use.

Also Read: I Stopped Uploading Private Documents to AI — This Free App Handles Them on My PC With Ollama | Anthropic Launches OSS Scanner, a Free AI Vulnerability Scanner for Open Source

Source: How-To Geek — read the full story there.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *