| name | prompt-compression |
| description | When you MUST feed a large blob into context (a long log, transcript, doc, or dataset dump), compress it to the salient parts first instead of pasting it whole. Use before including big inputs you cannot avoid. Distinct from context-warden (session discipline); this compresses one specific oversized input. Trigger with /prompt-compression or "compress this before adding", "summarize this blob for context", "this log is huge". |
| version | 0.1.0 |
| user-invocable | true |
| metadata | {"emoji":"🗜️"} |
prompt-compression
Sometimes you genuinely need a big external blob (a 5k-line log, a long transcript, a spec dump) in context. Pasting it whole is expensive and dilutes attention. Compress it to the parts that carry signal first.
Why this exists (evidence)
- Prompt compression (LLMLingua, Jiang et al., arXiv:2310.05736 and follow-ups) shows large prompts can be compressed multiple-x with little task-performance loss, cutting cost and latency, because most tokens in a big blob are low-information.
- It also fights context rot: fewer irrelevant tokens means the model attends to what matters.
When to use
- A SINGLE large input you cannot avoid including: long logs, stack traces, transcripts, large docs, data dumps, search results.
- Before pasting that blob into context or a sub-agent prompt.
- NOT a substitute for context-warden: that governs the whole SESSION (what to keep/drop over time); this compresses ONE oversized input on the way in.
The method
- Identify the signal the task needs from the blob: the error + its frames, the relevant section, the rows that matter, the decisions, the numbers.
- Extract, don't summarize loosely: keep exact identifiers (names, signatures, error strings, IDs, line refs), drop boilerplate, repetition, timestamps, banners, passing/no-op lines.
- Structure the residue: a short ordered extract or a small table, with a pointer back to the source (file:line / log range) so detail is recoverable on demand.
- State the compression: note what was dropped (e.g. "kept the 3 ERROR frames, dropped 4.8k INFO lines") so nothing looks hidden.
How to run it
- Logs: grep the error/levels you need, keep those frames + surrounding context, drop the rest.
- Transcripts/docs: extract the decisions/claims/sections relevant to the task; reference the rest by location.
- Heavy/automated: if an LLMLingua-style compressor or a summarizer tool is available, run it on the blob, then verify identifiers survived.
Composes with
context-warden: warden keeps the SESSION lean; prompt-compression shrinks a specific INPUT before it enters. Use together.retrieval-router: prefer retrieving only the needed slice over compressing the whole; compress when you truly must include a lot.run-cost: big blobs are top cost drivers; compress before paying for them repeatedly.
Honest limits
- Lossy by nature: aggressive compression can drop a detail that mattered. Keep exact identifiers and a pointer to the source so you can re-expand.
- For precise work, retrieving the exact slice (retrieval-router) beats compressing the whole. Compression is for when you cannot avoid the bulk.
- Cited ratios are from the papers' setups; measure your own loss.
