petermcneil/rlm

Overview

This skill uses recursive context processing to inspect and transform large codebases without loading everything into the model context. It treats the repository as an external environment, driving searches and focused sub-analyses with shell tools and small programmatic agents. The approach is optimized for multi-file refactors, large-scale analysis, and workflows where grep/find/uv are more effective than bulk context loading.

How this skill works

Initialize a REPL-style environment where the codebase is referenced symbolically (for example, as a CODEBASE variable). Use grep, find, sed/awk and small Python scripts (run via uv run) to locate, extract, and pre-process relevant slices of code. For each slice, create a focused prompt or sub-call to perform analysis or transformation, then aggregate and apply results incrementally with git-backed safety.

When to use it

• Working with repositories larger than the model's context window
• Multi-file refactoring, migrations, or large-scale code edits
• Cross-cutting analyses that require searching many files (security, dependency, or style audits)
• When iterative, slice-and-conquer processing is more reliable than loading full source
• Automated or repeatable transformations where scripting and logging are needed

Best practices

• Never load the entire codebase into the model; use symbolic handles for files or directories
• Use grep/find to narrow targets, then extract minimal focused context for each sub-call
• Run transformations incrementally and commit via git to enable safe rollbacks
• Store intermediate outputs in variables or files and log progress for debugging
• Limit recursion depth and add failure handling to avoid infinite loops

Example use cases

• Refactor a 10k+ line codebase by finding pattern matches and rewriting each file in focused sub-calls
• Perform a cross-file security scan by grepping for risky APIs, then symbolically analyzing matched functions
• Migrate configuration or API usage across many modules using scripted find + targeted prompts
• Generate per-file summaries or type stubs by extracting function definitions and invoking focused analysis
• Automate bulk lint fixes by locating offending patterns and applying tested edits file-by-file