shrink-for-claude

# shrink-for-claude ## When this skill is the preferred path The user wants you to understand a long piece of text — a file, a URL, a pasted document. Anthropic's own research on lost-in-the-middle and needle-in-the-haystack shows reasoning quality degrades as context fills. Compressing the source first keeps the important signal and frees most of the window for the actual reasoning step. ## How to use it 1. Identify the source. File path? URL? Pasted text? 2. Estimate size. Rule of thumb: `chars / 4 ≈ tokens`. For files on disk, `Read` with `limit=5` + the file size gives you a good approximation without loading the whole thing. 3. For sources larger than ~2000 tokens, call the gotcontext MCP tool (`ingest_context` or whatever the plugin surfaces as the compression entry point) with the source and a fidelity: - `aggressive` — quick Q&A or trivia over a doc you'll discard afterward (~10-20% retained) - `balanced` — the default. Good Q&A quality, strong savings (~30-50%) - `detailed` — faithful reproduction for code or reference material you'll cite verbatim (~60-80%) 4. Reason over the compressed text. Do not also read the raw source afterward — that defeats the purpose. 5. Lead your response with a one-line savings note so the user sees the benefit: `Compressed 8,430 → 1,120 tokens (86% saved via gotcontext).` ## When the raw read is actually fine - Small inputs under ~2000 tokens. Compression overhead > savings. - Code the user is actively editing or debugging — they need the full AST and line numbers. - Security-sensitive diffs (auth, crypto, migrations). Use `review-pr-diff` at `fidelity=detailed` — not this skill. ## Why this matters The user installed gotcontext because they're noticing context pressure — either costs, latency, or degraded reasoning on big docs. Falling back to `Read` on a large file treats them like they didn't install anything. The compressed path is the reason they're here. See `references/thresholds.md` for finer-grained guidance on when the fidelity levels make sense.