Overview

ClaudeShrink compresses large inputs using LLMLingua (gpt2) before you reason over them. This preserves semantic content while dramatically reducing token usage.

The compressor lives at: ~/.claude/skills/ClaudeShrink/scripts/compressor.py It runs inside an isolated venv at: ~/.claude/skills/ClaudeShrink/.venv

When to Use

• User pastes a large block of text, logs, or a document (>~8000 chars / ~2000 tokens)
• User asks to analyze, summarize, or reason over a large file on disk
• User's prompt is very long and would benefit from compression before reasoning
• User explicitly says "use ClaudeShrink" or "compress this"

Instructions

Follow these steps in order every time this skill is triggered:

1. Self-check: verify the environment is installed. Run:

   test -f ~/.claude/skills/ClaudeShrink/.venv/bin/python && echo "ready" || echo "not_installed"
   

   • If output is ready, proceed to step 2.
   • If output is not_installed, run the installer first:

     bash ~/.claude/skills/ClaudeShrink/install.sh
     

     If install.sh is missing (skill was added without cloning), fetch and run it:

     curl -fsSL https://raw.githubusercontent.com/g-akshay/ClaudeShrink/main/install.sh | bash
     

     Wait for it to complete, then proceed to step 2.

2. Identify the input source — is it a file path, raw pasted text, or a prompt?

3. Extract user intent — look at the user's request and derive a --question flag that captures what they care about. Examples:

   • "find all errors" → --question "What errors occurred?"
   • "summarize payment failures" → --question "What payment failures occurred?"
   • "keep all WARNING and ERROR lines" → --question "What warnings and errors occurred?"
   • No specific focus → omit --question (blind compression)

4. If it's a file on disk, run:

   ~/.claude/skills/ClaudeShrink/.venv/bin/python ~/.claude/skills/ClaudeShrink/scripts/compressor.py /absolute/path/to/file.txt --question "derived question here"
   

5. If it's raw pasted text or a prompt (no file on disk), write to a uniquely-named temp file, compress, then delete: Write the actual input content into the heredoc (do not write a placeholder string):

   TMP=$(mktemp /tmp/cs_input.XXXXXX.txt)
   cat > "$TMP" << 'EOF'
   [insert the full raw text content here]
   EOF
   ~/.claude/skills/ClaudeShrink/.venv/bin/python ~/.claude/skills/ClaudeShrink/scripts/compressor.py "$TMP" --question "derived question here"
   rm "$TMP"
   

6. Capture stdout — this is the compressed text. Ignore stderr (it contains stats for your reference).

7. If the compressor exits non-zero, warn the user ("ClaudeShrink compression failed — proceeding with raw input") and continue with the original uncompressed text.

8. Use only the compressed text (or raw text on failure) as your working context for the user's request.

9. Inform the user with a one-line note, e.g.:

   "Input compressed with ClaudeShrink (LLMLingua). Compression stats: [paste ratio from stderr if available]."

10. Proceed with the user's original request using the compressed context.

Output Format

• Do not show the raw compressed text to the user unless they ask for it.
• Respond to the user's original request (summarize, analyze, explain, etc.) as normal.
• Optionally append a brief compression note: original size, compressed token target, ratio.

Examples

Example 1 — Large log file with intent:

User: "Find all payment failures in this log: /var/log/app.log"

~/.claude/skills/ClaudeShrink/.venv/bin/python ~/.claude/skills/ClaudeShrink/scripts/compressor.py /var/log/app.log --question "What payment failures occurred?"

Then analyze the compressed output.

Example 2 — Pasted text with intent:

User: "Summarize the errors in this log" then pastes 800 lines.

TMP=$(mktemp /tmp/cs_input.XXXXXX.txt)
cat > "$TMP" << 'EOF'
[full pasted content]
EOF
~/.claude/skills/ClaudeShrink/.venv/bin/python ~/.claude/skills/ClaudeShrink/scripts/compressor.py "$TMP" --question "What errors occurred?"
rm "$TMP"

Example 3 — No specific focus:

User: "Compress this before you read it: [long prompt]"

Omit --question — blind compression applies.