recursive-decomposition
Agent skill for handling long-context tasks through recursive decomposition strategies based on RLM research (Zhang, Kraska, Khattab 2025)
pinned to #1780d46updated 2 weeks ago
Ask your AI client: “install plugins/recursive-decomposition”.
Requires the metahub MCP server installed in your client. Set up MCP.
mh install plugins/recursive-decompositionmetahub onboarded this repo on the author's behalf.
If you own github.com/massimodeluisa/recursive-decomposition-skill on GitHub, claim the listing to take over publishing. Your claim preserves the existing eval history and badges; only the curator label is replaced with verified-publisher on your next publish.
Stars
40
Last commit
2 weeks ago
Latest release
published
- #ai-agents
- #claude-code-plugin
- #claude-code-skill
- #claude-code-skills
- #claude-skills
- #llm
- #long-context
- #recursive-decomposition
- #rlm
Evaluation report
WarningsAutomated checks the publisher passed at publish time — structure, docs, safety, and whether the artifact behaves as claimed.1780d46· 2 weeks ago
Kind-specific
31Plugin: manifest fields complete
Plugin: bundled artifacts presentwarn
Couldn't find a skills/ subdir or an MCP server config inside the plugin
A plugin is most useful when it bundles ≥ 1 skill or an MCP server.
Plugin: bundle shape
empty bundle
Plugin: manifest location
manifest at plugin.json
Release history
1- releasecurrent1780d46warn2 weeks ago
Contents
A Claude Code plugin for handling long-context tasks through recursive decomposition strategies based on Recursive Language Models (RLM) research by Zhang, Kraska, and Khattab (2025).
What This Plugin Does
This plugin enables Claude to handle tasks that exceed comfortable context limits (typically 30k+ tokens or 10+ files) by:
- Filtering before deep analysis - Narrowing search space systematically
- Strategic chunking - Partitioning inputs for parallel processing
- Recursive sub-agents - Spawning independent analysis tasks
- Answer verification - Re-checking synthesized results on focused windows
- Programmatic synthesis - Aggregating results without context overflow
Real-world impact: Handles inputs up to 2 orders of magnitude beyond normal context limits, with benchmarks showing 21-58% accuracy improvements on long-context tasks.
When to Use
The skill automatically activates when Claude detects:
- Tasks involving 10+ files or 50k+ tokens
- Phrases like "analyze all files", "process this large document", "aggregate information from", "search across the codebase"
- Codebase-wide pattern analysis
- Multi-document information extraction
- Multi-hop questions requiring scattered evidence
Skills Included
- recursive-decomposition - Core strategies for programmatic decomposition, chunking, filtering, and recursive self-invocation
Installation
From Marketplace
# Add this marketplace
/plugin marketplace add massimodeluisa/recursive-decomposition-skill
# Install the plugin
/plugin install recursive-decomposition
Manual Installation
Copy this plugin directory to your Claude Code plugins directory:
cp -r plugins/recursive-decomposition ~/.claude/plugins/
Usage Examples
"Find all error handling patterns across the entire codebase"
"Summarize all TODO comments and categorize by priority"
"What API endpoints are defined across all route files?"
"Analyze security vulnerabilities in all Python files"
The plugin handles the decomposition automatically - no special syntax required.
Plugin Structure
recursive-decomposition/
├── .claude-plugin/
│ └── plugin.json # Plugin manifest
└── skills/
└── recursive-decomposition/
├── SKILL.md # Core decomposition strategies
└── references/
├── rlm-strategies.md
├── cost-analysis.md
├── codebase-analysis.md
└── document-aggregation.md
Research Foundation
Based on "Recursive Language Models" (Zhang, Kraska, Khattab, 2025):
- arXiv: 2512.24601
- Enables context scaling from 2^14 to 2^18 tokens (16x)
- ~3x more cost-effective than summarization baselines
License
MIT - See LICENSE
Author
Massimo De Luisa (@massimodeluisa)
Reviews
No reviews yet. Be the first.
Related
espalier-engineering
Train your AI coders the way you'd train a vine — discover your codebase's actual patterns, then encode them as rules, skills, agents, hooks, and a guided pipeline so generated code lands inside your conventions on the first try, not the fifth
superloopy
Lightweight loop harness with strict evidence gates — Claude Code edition.
beagle-ai
Pydantic AI, LangGraph, DeepAgents, and Vercel AI SDK skills for building and reviewing AI applications.
mh install plugins/recursive-decomposition