Usage: USAGE.md — install, CLI, config, workflows.
Beginner: BEGINNER.md · All docs: docs/README.md
Open-source toolkit for token-efficient LLM agent skills, grounded in three research papers:
| # | Paper | What it does | How you run it |
|---|---|---|---|
| 1 | SkillReducer (Gao et al.) | Compress skill description + body (+ optional scripts) |
skillreducer reduce / agent |
| 2 | TSCG (Sakizli) | Compress MCP / tool JSON schemas | skillreducer reduce … --tscg |
| 3 | SkillRevise (Liu et al.) | Revise skill quality from execution traces | skillreducer revise (separate command) |
Optional quality pass → SkillRevise → skillreducer revise …
Skill text tokens → SkillReducer St.1–3 → lean SKILL.md + refs + scripts/
Tool / MCP schemas → TSCG → lean mcp_manifest.tscg.*
Works with any agent platform that uses the standard SKILL.md + YAML frontmatter convention (Claude Code, Windsurf, OpenCode, SkillHub, GitHub community skills, and similar).
pip install -e .
skillreducer audit data/pdf-processing
skillreducer reduce data/pdf-processing --no-llmFull install, flags, config, TSCG, and revise: USAGE.md.
USAGE.md ← top-level usage details (start here to run tools)
BEGINNER.md ← beginner hub
docs/
README.md ← documentation index
OVERVIEW.md ← how the three papers fit
PAPERS.md ← paper index
REDUCTION_FLOW.md ← flow + worked example
skillreducer/ ← PAPER · BEGINNER · USAGE
tscg/ ← PAPER · BEGINNER · USAGE
skillrevise/ ← PAPER · BEGINNER · USAGE · DEVELOPER
CITATION.md
| Resource | Link |
|---|---|
| Usage (top-level) | USAGE.md |
| Docs index | docs/README.md |
| SkillReducer | docs/skillreducer/ |
| TSCG | docs/tscg/ |
| SkillRevise | docs/skillrevise/ |
| Papers overview | docs/OVERVIEW.md · docs/PAPERS.md |
| Citations | CITATION.md |
| SkillReducer arXiv | 2603.29919 · PDF |
| TSCG arXiv | 2605.04107 |
| SkillRevise arXiv | 2606.01139 |
Compress routing description, restructure the body with progressive disclosure, optionally extract scripts.
Docs: PAPER · BEGINNER · USAGE · stage1 · stage2 · stage3
Optional --tscg after skill reduction; you provide tools JSON.
Docs: PAPER · BEGINNER · USAGE
Separate revise command; vendored under src/skillrevise/.
Docs: PAPER · BEGINNER · USAGE · DEVELOPER
flowchart TD
In[SKILL.md + optional refs] --> S1[Stage 1: compress / generate description]
S1 --> S2[Stage 2: classify body + progressive disclosure]
S2 --> S3[Stage 3: selective Python / bash → scripts/]
S3 --> Out[optimized/skill/: SKILL.md + refs + scripts/]
Out --> OptTSCG[Optional: --tscg on tools.json]
Default reduce / agent runs Stage 1 → 2 → 3. Use --stage N for a single stage.
Commands and config: USAGE.md.
- Never modifies skills in-place by default; output goes to
--output.
pytest
ruff check skillreducer testsCite the paper authors for methods and numbers — not this repository alone.
CITATION.md · docs/PAPERS.md
MIT — see LICENSE. The research papers are © their authors; this repo is an independent implementation.