Compress Text

compress-text-skill

An Agent Skill that compresses and restructures any text using MECE grouping and the Pyramid Principle. It generates multiple candidate groupings, scores each on size reduction and meaning-fidelity, and ships the best one — preserving every instruction, key point, reference, and example.

It compresses through four complementary lenses:

• Semantic — merge paraphrases and entailments into the single most specific statement (meaning-level dedup, even when wordings differ).
• Statistical — scan for repeated n-grams, near-duplicate sentences, and low-density filler to target redundancy objectively and report how much was removed.
• Structural (key-point tree) — when the input spans multiple related documents or sections, build one whole-context key-point tree: hoist shared points to their common ancestor and collapse cross-document duplicates so each point is stated once, preserving the union of distinct points.
• Supersession (pivots) — in transcripts, dictation, drafts, and chat, drop self-corrections ("actually", "scratch that", "I mean", "instead") together with the passages they supersede, keeping only the final intent.

Supports two output modes:

• dense (LLM-oriented): maximize information per token for prompts and context.
• readable (human-oriented): keep natural, fluent prose for people.

It also preserves template tags verbatim — auto-detected engine tags (Handlebars/Mustache/Jinja/Liquid/shell/ERB) and explicit <!-- keep --> / <!-- drop --> markers — and can emit the result into a user-supplied output template with named <!-- slot:NAME --> slots.

Repository layout

compress-text-skill/
├── .claude-plugin/
│   ├── plugin.json          # Claude Code plugin manifest
│   └── marketplace.json     # Claude Code marketplace catalog (this repo)
├── .github/
│   └── workflows/ci.yml     # CI: validate + run evals + build bundle
├── skills/
│   └── compress-text/
│       └── SKILL.md          # The skill
├── evals/                   # Evaluation suite (see evals/README.md)
├── scripts/
│   ├── build-skill-bundle.sh
│   └── validate.sh
├── CHANGELOG.md
├── LICENSE
└── README.md

This repo is both a Claude Code plugin (root contains skills/) and a marketplace that lists that plugin.

Install via Claude Code plugin marketplace

# Add this repo as a marketplace
/plugin marketplace add MartinForReal/compress-text-skill

# Install the plugin
/plugin install compress-text@martinforreal-skills

Once installed, ask Claude to "compress this text" (or invoke the skill directly) and it will apply the MECE + Pyramid workflow.

To validate the manifests locally before publishing:

claude plugin validate . --strict

Publish the skill bundle (claude.ai / Skills API / Foundry)

Build a distributable .zip bundle:

./scripts/build-skill-bundle.sh
# -> dist/compress-text-skill.zip  (contains compress-text/SKILL.md)

Then upload the zip wherever Agent Skills are accepted:

• claude.ai — Settings → Capabilities → Skills → Upload skill
• Claude Skills API — POST the bundle to the Skills endpoint
• Claude Platform on AWS / Microsoft Foundry — upload via the Skills API

Usage examples

• "Compress this for an LLM prompt: …" (dense mode)
• "Shorten these meeting notes but keep them readable" (readable mode)
• "Deduplicate and restructure this doc, show me the candidate comparison"
• "Clean up this dictated note: drop the parts where I changed my mind and keep the final decisions" (pivot/supersession)
• "Compress this and tell me how much redundancy you removed" (statistical reporting)
• "Compress this email template but keep the {{ tags }} untouched"
• "Compress #file:notes.md and report the token reduction"

Development

Validate everything (manifests, skill structure, eval datasets) before publishing:

./scripts/validate.sh

Run just the evaluation suite (offline static checks, no model required):

python3 evals/run_evals.py            # structure + dataset integrity
python3 evals/run_functional.py --selftest   # deterministic behaviour checks (30/30 cases)

The suite also includes a labelled triggering dataset, functional rubric cases for model-judged evaluation, and a SkillOpt held-out validation loop (train/val split + evals/OPTIMIZATION_LOG.md) that gates skill edits on generalization — see evals/README.md. CI runs the validator and builds the bundle on every push and pull request.

Changelog

See CHANGELOG.md.

License

MIT