rlm

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, code) that won't fit comfortably in a single prompt.
• You need to iteratively inspect, search, chunk, and extract information from that context.
• You can delegate chunk-level analysis to the rlm-subcall subagent.

Mental model

• This Claude Code conversation = the root LM.
• Persistent Python REPL (rlm_repl.py) = the external memory / environment.
• Subagent rlm-subcall = the sub-LM (equivalent to llm_query in the RLM paper).

Inputs

Accept these patterns from $ARGUMENTS:

• context=<path> (required): path to the large context file.
• query=<question> (required): what to answer.
• Optional: chunk_chars=<int> (default 150000) and overlap_chars=<int> (default 0).

If arguments are missing, ask for the context file path and query.

Step-by-step procedure

1. Initialise 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

   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(f'Total: {len(content):,} chars')"
   

3. Choose a chunking strategy

   • Semantic: split on markdown headers, JSON objects, log timestamps, etc.
   • Fallback: fixed character chunks (~150K chars each).

4. Write chunks to files (so rlm-subcall can read them)

   python3 .claude/skills/rlm/scripts/rlm_repl.py exec <<'PY'
   paths = write_chunks('.claude/rlm_state/chunks', size=150000, overlap=0)
   print(f'{len(paths)} chunks written')
   PY
   

5. Subcall loop — for each chunk, invoke rlm-subcall with:

   • The user query
   • The chunk file path
   • Any specific extraction instructions Collect each subagent result into buffers:

   python3 .claude/skills/rlm/scripts/rlm_repl.py exec -c "add_buffer('<result from subagent>')"
   

6. Synthesise — combine buffers in the main conversation to produce the final answer. Optionally pass all buffers to rlm-subcall once more for a coherent draft.

Guardrails

• Do not paste large raw chunks into the main conversation.
• Use grep/peek to locate relevant excerpts; quote only what you need.
• Subagents cannot spawn other subagents — all orchestration stays here.
• Keep scratch files under .claude/rlm_state/.

Alternative approaches for this context

For reproducible, non-interactive pipelines consider the Docker-based alternatives in this repo:

• examples/agentic/long_context/rlm_needle/ — needle-in-haystack demo (any LLM API)
• examples/agentic/long_context/rlm_doc_qa/ — document QA (any LLM API)