Everyone optimizes what they type into Claude.
Nobody optimizes what they paste.
But developers paste constantly. GitHub READMEs. Research papers. API docs. Jira tickets. Confluence pages. Slack threads.
Every time you copy a webpage, your clipboard picks up everything. The navigation. The footer. The boilerplate. The cookie banners. The share buttons rendered as plain text.
You wanted the content. Claude got the garbage too. And you paid for every token of it.
The numbers
Real token counts from live Claude Code sessions:
| Content | Before | After | Saved |
|---|---|---|---|
| GitHub README (trooper) | 4,600 | 1,200 | 74% |
| Research paper (arXiv) | 4,800 | 1,900 | 60% |
| GitHub README (caveman) | 1,800 | 600 | 67% |
| API documentation | 105 | 35 | 67% |
| tokslayer's own README | 800 | 170 | 79% |
| Average | 67% |
Where the savings actually land
First turn input tokens are paid in full. The saving is on two things:
Output tokens. Claude responds to the compressed version, so answers are shorter and more focused.
Session history. The compressed version stays in context, not the bloated original. Every subsequent turn in that session carries 3,400 fewer tokens of history. Long sessions with multiple pastes, this compounds hard.
Note: A subagent-based approach was considered but doesn't fully solve this .It can't shrink the user's original pasted message in that turn. A write-path
proxy is the correct fix, rewriting the outgoing request on every turn
including history resends. That's the roadmap direction.
What's actually happening when you paste
When you copy a GitHub page you get the content plus:
- "Skip to content" navigation
- Repo tabs (Code, Issues, Pull requests, Actions...)
- Breadcrumbs and branch selectors
- Footer links
- Share buttons rendered as text
- License metadata
None of that is the README. All of it hits Claude's context window. All of it costs tokens.
The fix
A Claude Code skill that sits between your clipboard and Claude. Detects pasted content. Strips the noise. Sends only the signal.
No proxy. No server. No MCP. No configuration. One file. Drop it in. Restart Claude Code. Done.
Receipt on every paste:
ORIGINAL: "Skip to content shouvik12 trooper Repository..." (~4,600 tokens)
OPTIMIZED: "Trooper: LLM proxy. Local-first. Ollama default..." (~1,200 tokens)
SAVED: ~3,400 tokens (74%)
What gets stripped: nav chrome, footers, filler phrases, redundant sentences, marketing boilerplate.
What stays: headings, code blocks, API signatures, URLs, numbers, technical terms, proper nouns.
Install
curl -fsSL https://raw.githubusercontent.com/shouvik12/tokslayer/main/install.sh | bash
Restart Claude Code. Works on every paste automatically from that point on.
The meta test
Ran tokslayer's own README through itself and typed "summarize":
ORIGINAL: "tokslayer, Slays tokens before they reach Claude..." (~800 tokens)
OPTIMIZED: "Tokslayer: Claude Code skill. Compresses pasted content..." (~170 tokens)
SAVED: ~630 tokens (79%)
A tool that eats its own cooking.
Where it fits
This covers the input side. Pair it with caveman for output compression
(65% reduction on Claude responses).
Together: lean input, lean output
For further actions, you may consider blocking this person and/or reporting abuse
