Recursive Decomposition Skill

What it is

Recursive Decomposition Skill is Massimo De Luisa’s skill for long-context jobs, such as a whole codebase, a folder of documents or a pile of PDFs, even when they fit the context window. It makes Claude size the input first, search before it reads, and give each batch of files to its own sub-agent. Claude then checks the merged answer against the sources. It rests on the Recursive Language Models paper, which names the problem context rot.

1. You ask about a whole repo, a document folder or PDFs
2. Claude sizes the input and searches before it reads
3. Claude gives each batch to one sub-agent, one level deep
4. Claude merges the answers and spot-checks the sources

Who it suits

People who point Claude at something too large, dense or scattered to trust to one read. Think of an inherited codebase, a docs folder full of PRDs or meeting notes, or a stack of PDFs. It is not for one file or one function. For those, the skill tells Claude to read directly.

Where to use it

The README names Claude Code, Codex, Cursor and compatible agents, and it ships a Claude Code plugin.

• Claude Code, all your projects: the skills CLI with -g, or the plugin
• Claude Code, one project: a copy of the skill folder in that project’s .claude/skills/
• Codex, Cursor and other agents: the skills CLI with -a, or a copy of the folder in the agent’s skills folder

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with every route the README gives.

1. Install the skill with one command. The line below uses the skills CLI. Add -g for a user-level install, or -a claude-code to aim it at one agent.
2. Start your agent, or restart it. Claude Code’s docs say an open session sees a new skill at once, unless its skills folder was made after the session started. Then run /reload-skills or restart.
3. Run /recursive-decomposition on a task. Give it a path or a question, and Claude does the sizing, searching and batching.

npx skills add massimodeluisa/recursive-decomposition-skill

Claude Code can also take it as a plugin. Run these two lines in a terminal, not inside a session. The marketplace registers as recursive-decomposition-skill, so the install reads the plugin’s name, @ and that name. Claude Code’s docs say a plugin skill then runs as /recursive-decomposition:recursive-decomposition.

claude plugin marketplace add massimodeluisa/recursive-decomposition-skill
claude plugin install recursive-decomposition@recursive-decomposition-skill

To copy it by hand, take the folder skills/recursive-decomposition from the repository into ~/.claude/skills/ and restart the agent. The README gives no command for that. Pick one route, because Claude Code’s docs say a plugin skill and a copied skill both load, so you would see it twice.

How to use it

Run /recursive-decomposition with a path or a question, or ask in plain words and let the skill load by itself. These are the README’s own prompts.

Analyze error handling patterns across this entire codebase
Find all TODO comments in the project and categorize by priority
What API endpoints are defined across all route files?
Summarize the key decisions from all meeting notes in docs/
Find security issues across all Python files

Claude does the six steps below, and you read the answer.

Example

You inherit a large TypeScript project and want to know how it handles errors everywhere, not in one file.

/recursive-decomposition Analyze error handling patterns across this entire codebase

You get one report of the error types by area, what already agrees, what does not and what to change, with file references.

For newcomers and experts

If you are new. A skill is a folder with a SKILL.md file that your agent reads when your task fits its description. This one is for big jobs only. Try it on a docs folder first, and ask for one list that covers every file. Context rot means answers get worse as the input grows, even inside a large window.

For experts. SKILL.md states hard rules. Claude must size the input before it reads and search a folder instead of listing its whole tree. A file over 2,000 lines or 50 KB is read by line range. A PDF over 100 pages or 30 MB is read by metadata or after a split. Claude may load no more than 5 files into the main context without a written batch plan.

Each sub-agent gets its own brief with the files, the question and the output schema. Depth stays at 1, and the batch count is written before launch. Claude never raises depth because a merge looks thin. It re-reads the disagreement instead. In the paper GPT-5, run as a recursive model, scored 58.0 F1 at depth 1 and 76.0 at depth 3 on OOLONG-Pairs. The skill’s notes add that another model can score worse when it goes deeper, so the skill stays at depth 1.

For PDFs, Claude first converts the file to markdown, then searches the markdown. Claude runs this line, with your file in place of FILE.

npx -y @firecrawl/anydoc FILE -o .firecrawl/out.md

Scanned pages make anydoc exit with code 3. Claude then reruns it with --ocr hosted or calls Firecrawl’s cloud firecrawl parse. SKILL.md gives that cloud step about 1 credit per PDF page and a 50 MB cap.

Good to know

It is one skill with no script, and it spends extra tokens on coordination.

• Cost: The README says the paper’s runs cost about 3x less than summarising. The skill’s own cost page lists $0.99 against $0.57 for compaction on one task. Treat cost as unsettled
• Small jobs: for one file or a quick lookup, reading directly is faster
• Paper numbers: on multi-hop questions over 6 to 11 million tokens, SKILL.md gives 91.3% for the method against 70.5% for compaction. Inputs that size are far larger than most projects, and the score is no promise for your files
• Maker’s check: on 131 public PDFs, plain pdftotext was faster but got no title from the largest file, a scan. The skill was slower because it ran OCR on that file
• Privacy: your files and the sub-agents’ briefs go to your agent’s model as in any other work. A scanned PDF sent to the cloud parse goes to Firecrawl
• Updates: the README gives no update or removal step
• Maker: Massimo De Luisa, MIT licensed. An independent project, not affiliated with the paper’s authors, the university MIT, tccao or Firecrawl. Mindket lists the skill and is not the maker