Skills Guide

Workflow RLM (Modèle de Langage Récursif)

VérifiéPrudence

Permet de traiter de très grands contextes en déléguant l'analyse par morceaux à un sous-agent, avec un REPL Python persistant.

Spar Skills Guide Bot
Data & IAAvancé
1026/07/2026
Claude Code
#recursive-language-model#long-context#subagent#chunking#data-extraction

Recommandé pour

Claude CodeData scientistData & analyse

Notre avis

Ce skill exécute une boucle de modèle de langage récursif pour traiter de grands contextes en les segmentant et en déléguant l'analyse à un sous-agent.

Points forts

  • Gère des fichiers dépassant la limite de contexte du chat
  • Utilise un REPL Python pour une manipulation efficace des données
  • Permet un chunking sémantique adaptatif
  • Orchestre des sous-agents pour une extraction parallélisée

Limites

  • Nécessite une configuration manuelle du REPL et des chunks
  • Les sous-agents ne peuvent pas créer d'autres sous-agents
  • La synthèse finale repose sur le jugement de l'agent principal
Quand l'utiliser

Utilisez ce skill lorsque vous devez extraire des informations d'un document très volumineux (logs, transcriptions, pages web) qui ne tient pas dans le contexte de la conversation.

Quand l'éviter

Ne l'utilisez pas pour des documents de taille modérée qui tiennent dans le contexte chat, ou lorsque la tâche ne nécessite pas de découpage ni d'analyse multi-étapes.

Analyse de sécurité

Prudence
Score qualité90/100

The skill involves running Bash commands and writing files as part of a legitimate long-context processing workflow. No inherently destructive or exfiltrating commands are present, and the operations are scoped to a dedicated state directory. However, the use of Bash and dynamic Python execution warrants a caution label.

Points d'attention
  • •The skill uses Bash to execute Python scripts, which could be exploited if arguments are unsanitized, though typical usage assumes safe inputs.
  • •It writes files to .claude/rlm_state/ and chunks, but this is within expected boundaries for the workflow.

Exemples

Analyze a large log file
Use the rlm skill to analyze this large log file: /var/log/application.log. The query is: find all occurrences of ERROR in the last 24 hours and summarize the most common error types.
Extract facts from a lengthy transcript
Run rlm with context=/path/to/transcript.txt and query='Extract all mentions of project deadlines and responsible persons, output as JSON.'
Summarize a book-length document
I have a 500-page technical document at ./docs/guide.pdf. Use rlm to chunk it and provide a structured summary of each chapter.

name: rlm description: Run a Recursive Language Model-style loop for long-context tasks. Uses a persistent local Python REPL and an rlm-subcall subagent as the sub-LLM (llm_query). allowed-tools:

  • Read
  • Write
  • Edit
  • Grep
  • Glob
  • Bash

rlm (Recursive Language Model workflow)

Use this Skill when:

  • The user provides (or references) a very large context file (docs, logs, transcripts, scraped webpages) that won't fit comfortably in chat context.
  • You need to iteratively inspect, search, chunk, and extract information from that context.
  • You can delegate chunk-level analysis to a subagent.

Mental model

  • Main Claude Code conversation = the root LM.
  • Persistent Python REPL (rlm_repl.py) = the external environment.
  • Subagent rlm-subcall = the sub-LM used like llm_query.

How to run

Inputs

This Skill reads $ARGUMENTS. Accept these patterns:

  • context=<path> (required): path to the file containing the large context.
  • query=<question> (required): what the user wants.
  • Optional: chunk_chars=<int> (default ~200000) and overlap_chars=<int> (default 0).

If the user didn't supply arguments, ask for:

  1. the context file path, and
  2. the query.

Step-by-step procedure

  1. Initialise the REPL state

    python3 .claude/skills/rlm/scripts/rlm_repl.py init <context_path>
python3 .claude/skills/rlm/scripts/rlm_repl.py status

  2. Scout the context quickly

    python3 .claude/skills/rlm/scripts/rlm_repl.py exec -c "print(peek(0, 3000))"
python3 .claude/skills/rlm/scripts/rlm_repl.py exec -c "print(peek(len(content)-3000, len(content)))"

  3. Choose a chunking strategy

    • Prefer semantic chunking if the format is clear (markdown headings, JSON objects, log timestamps).
    • Otherwise, chunk by characters (size around chunk_chars, optional overlap).

  4. Materialise chunks as files (so subagents can read them)

    python3 .claude/skills/rlm/scripts/rlm_repl.py exec <<'PY'
paths = write_chunks('.claude/rlm_state/chunks', size=200000, overlap=0)
print(len(paths))
print(paths[:5])
PY

  5. Subcall loop (delegate to rlm-subcall)

    • For each chunk file, invoke the rlm-subcall subagent with:
      • the user query,
      • the chunk file path,
      • and any specific extraction instructions.
    • Keep subagent outputs compact and structured (JSON preferred).
    • Append each subagent result to buffers (either manually in chat, or by pasting into a REPL add_buffer(...) call).

  6. Synthesis

    • Once enough evidence is collected, synthesise the final answer in the main conversation.
    • Optionally ask rlm-subcall once more to merge the collected buffers into a coherent draft.

Guardrails

  • Do not paste large raw chunks into the main chat context.
  • Use the REPL to locate exact excerpts; quote only what you need.
  • Subagents cannot spawn other subagents. Any orchestration stays in the main conversation.
  • Keep scratch/state files under .claude/rlm_state/.
Skills similaires

Ingénierie de Prompts

Data & IA

Bonnes pratiques et templates de prompt engineering pour maximiser les résultats IA.

claudeCursorWindsurf+1beginner
289
78
868
Admin

Visualisation de Données

Data & IA

Génère des visualisations de données et graphiques adaptés à vos données.

claudeCursorWindsurfintermediate
198
56
703
Admin

Architecture RAG

Data & IA

Guide de configuration d'architectures RAG (Retrieval-Augmented Generation).

claudeCursorWindsurfadvanced
167
51
667
Admin