Compresr Context Compressor

Workflow

When this skill triggers, follow these steps in order.

Step 1 — Receive the long context

Accept the text to compress. This can be:

• A pasted document (design doc, chat transcript, code file, meeting notes, etc.)
• A file path the user provides
• Any large blob the user wants to fit into a tighter context window

If the user provides a file path, confirm it exists before proceeding.

Step 2 — Run the compressor

From the skill root directory, run:

node scripts/compress.mjs <inputFile> <outputFile>

The engine applies four deterministic passes in order:

1. Paragraph dedup — Finds paragraphs (blank-line-delimited blocks) that appear more than once and replaces every repeat with a _(repeated Nx, deduped)_ annotation on the first occurrence.
2. Boilerplate stripping — Removes lines matching known low-salience patterns: Lorem ipsum, legal confidentiality notices, copyright lines, "for internal use only" footers, and similar content.
3. Line dedup — Removes exact duplicate non-blank lines within the remaining text.
4. Blank-line collapse — Collapses runs of two or more consecutive blank lines into a single blank line, and strips trailing whitespace from each line.

The compressed text is written to the output file. Token estimates are printed to stderr for inspection.

Step 3 — Compute before/after token estimates

The skill uses scripts/count-tokens.mjs to estimate token counts before and after compression. The estimator uses a word + punctuation heuristic (~4 chars per alphanumeric run, 1 token per punctuation character). It is an approximation — see the caveat note below.

node scripts/count-tokens.mjs <file>

Step 4 — Report results to the user

Report the following:

• Tokens before (estimated)
• Tokens after (estimated)
• Percentage reduction
• The compressed context (inline or as a file reference)
• The caveat that counts are estimates

Output format

Compression complete.

  Tokens before: ~<N> (estimated)
  Tokens after:  ~<M> (estimated)
  Reduction:     <X>%

Note: token counts are heuristic estimates (word+punctuation splitter).
For exact BPE counts, use tiktoken (OpenAI) or LLMLingua (Microsoft).

Compressed context:
---
<compressed text here>
---

If the reduction is below 5%, tell the user the context is already compact and compression had minimal impact.

Example

See examples/input.md for a 200+ line design document with intentional redundancy (duplicated sections, repeated boilerplate, blank-line runs).

Run the example yourself:

cd skills/compresr-context-compressor
bash examples/run.sh

The output is written to examples/output.md with the full before/after report and the compressed context in a fenced block.

Caveats

• Lossy-but-safe: Compression removes genuinely redundant content. It does not summarize or paraphrase — unique content is always preserved.
• Estimate caveat: Token counts are heuristic estimates, not real BPE counts. For production use, replace count-tokens.mjs with tiktoken or LLMLingua.
• Boilerplate is deterministic: The boilerplate patterns are regex-based and documented in the source. If a pattern incorrectly strips content, remove or adjust it in scripts/compress.mjs.