ClaudeShrink - Large Text Compression

Description

--- name: claudeshrink version: 1.0.0 author: Akshay Gundewar description: > Automatically compress large natural text or log files before processing. Trigger when the user pastes massive text blobs, or asks to analyze a large file (logs, docs, transcripts), or provides a prompt that is too large for the context window. DO NOT trigger on source code files or structural data (JSON, XML). tags: - compression - tokens - context-window - llmlingua - skills - ai-tool - claude-code - prompt-compression requires: - python3 - git allowed-tools: - Bash --- # Overview ClaudeShrink compresses large inputs using [LLMLingua](https://github.com/microsoft/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: ```bash 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 bash ~/.claude/skills/ClaudeShrink/install.sh ``` If `install.sh` is missing (skill was added without cloning), fetch and run it: ```bash 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. **If it's a file on disk**, run: ```bash ~/.claude/skills/ClaudeShrink/.venv/bin/python ~/.claude/skills/ClaudeShrink/scripts/compressor.py /absolute/path/to/file.txt ``` 4. **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): ```bash 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 "$TMP" rm "$TMP" ``` 5. **Capture stdout** — this is the compressed text. Ignore stderr (it contains stats for your reference). 6. **If the compressor exits non-zero**, warn the user ("ClaudeShrink compression failed — proceeding with raw input") and continue with the original uncompressed text. 7. **Use only the compressed text** (or raw text on failure) as your working context for the user's request. 8. **Inform the user** with a one-line note, e.g.: > "Input compressed with ClaudeShrink (LLMLingua). Compression stats: [paste ratio from stderr if available]." 9. **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:** > User: "Analyze this error log: /var/log/app.log" ```bash ~/.claude/skills/ClaudeShrink/.venv/bin/python ~/.claude/skills/ClaudeShrink/scripts/compressor.py /var/log/app.log ``` Then analyze the compressed output. **Example 2 — Pasted text:** > User pastes 800 lines of documentation inline. Write it to a `mktemp`-generated path, compress, delete, analyze. **Example 3 — Explicit trigger:** > User: "Use ClaudeShrink on this prompt before answering: [long prompt]" Same as Example 2.

Security Status