RLM Processing Skill

Expert guidance for using Recursive Language Model (RLM) patterns to analyze large codebases efficiently.

Core Concept

RLM treats large context as an external environment rather than loading it directly into the LLM's context window. The LLM:

1. Loads files as variables in a REPL
2. Writes code to chunk, filter, and process
3. Uses llm_query() for semantic sub-tasks
4. Returns condensed, relevant results

When to Use

Trigger Conditions

• Token Threshold: Context exceeds 50K tokens
• File Count: Story requires understanding >5 interconnected files
• Pattern Discovery: Need to find where something happens in unfamiliar code
• Cross-Cutting Analysis: Auth, logging, error handling across modules

Activation Examples

Scenario: "Add rate limiting to all API endpoints"
- Need to find: All endpoint definitions
- Files involved: Likely 15+ route files
- Context size: ~100K tokens
- Decision: USE RLM

Scenario: "Fix typo in login button"
- Location: Known (src/components/Login.tsx)
- Files involved: 1
- Context size: ~2K tokens
- Decision: SKIP RLM

REPL Environment

Available Variables

context       # Dict: {file_path: file_content}
total_chars   # Total characters loaded
total_tokens  # Estimated tokens (~chars/4)

Available Functions

llm_query(prompt)     # Query sub-LLM for semantic tasks
print(text)           # Output (truncated to 30K chars)
FINAL(answer)         # Return text answer
FINAL_VAR(var_name)   # Return variable content

Chunking Strategies

By File Type

TypeScript/JavaScript:

# Split by exports/functions/classes
pattern = r'export\s+(?:const|function|class|interface|type)\s+\w+'

Python:

# Split by def/class
pattern = r'^(?:def|class)\s+\w+'

Markdown:

# Split by headers
pattern = r'^#{1,3}\s+.+'

By Size

def chunk_by_size(text, max_chars=5000):
    """Safe fallback for any file type"""
    chunks = []
    current = ""
    for line in text.split('\n'):
        if len(current) + len(line) > max_chars:
            chunks.append(current)
            current = line
        else:
            current += '\n' + line
    if current:
        chunks.append(current)
    return chunks

When to Use llm_query() vs Code

Use Code When:

Use llm_query() When:

Common Patterns

Find All Endpoints

endpoints = []
for path, content in context.items():
    if not path.endswith(('.ts', '.js')):
        continue
    # Express/Koa style
    rest = re.findall(r'\.(get|post|put|delete|patch)\([\'"]([^\'"]+)', content)
    # Next.js style
    if '/pages/api/' in path or '/app/api/' in path:
        endpoints.append((path, 'route', path))
    endpoints.extend([(path, m, r) for m, r in rest])

Map Authentication Flow

auth_files = [p for p in context.keys()
              if any(x in p.lower() for x in ['auth', 'login', 'session'])]

auth_map = {}
for path in auth_files[:5]:
    role = llm_query(f"What role does this play in auth? (1 sentence)\n\n{context[path][:2500]}")
    auth_map[path] = role

Find Error Handling

error_patterns = []
for path, content in context.items():
    if re.search(r'(try\s*\{|\.catch\(|catch\s*\(|throw\s+new)', content):
        # Count error handling instances
        count = len(re.findall(r'catch', content))
        error_patterns.append((path, count))

# Sort by error handling density
error_patterns.sort(key=lambda x: -x[1])

Identify Test Files

test_files = [p for p in context.keys()
              if re.search(r'(\.test\.|\.spec\.|__tests__)', p)]

# Map tests to source files
test_mapping = {}
for test_path in test_files:
    # Extract what's being tested
    source = re.sub(r'\.(test|spec)', '', test_path)
    source = re.sub(r'__tests__/', '', source)
    test_mapping[test_path] = source

Cost Optimization

Budget llm_query() Calls

MAX_SUB_CALLS = 20
sub_calls_made = 0

def budget_query(prompt):
    global sub_calls_made
    if sub_calls_made >= MAX_SUB_CALLS:
        return "[BUDGET EXCEEDED - skipping]"
    sub_calls_made += 1
    return llm_query(prompt)

Batch Similar Queries

# Instead of 10 separate calls:
# BAD: for f in files: llm_query(f"Analyze {f}")

# Batch into one:
batch = "\n---\n".join([f"File: {f}\n{context[f][:1000]}" for f in files[:5]])
analysis = llm_query(f"For each file below, state its purpose (1 line each):\n\n{batch}")

Filter Before Querying

# Don't query all files - filter first
relevant = [p for p in context.keys()
            if re.search(r'user|auth|login', context[p].lower())]

# Then query only relevant files
for path in relevant[:10]:
    # Now llm_query is worth the cost
    ...

Output Format

Always return structured JSON:

{
  "analysis": "Summary of findings",
  "relevant_files": ["src/auth/index.ts", "src/middleware/auth.ts"],
  "code_patterns": [
    "Pattern: JWT tokens stored in httpOnly cookies",
    "Pattern: Auth middleware applied via app.use() in src/app.ts"
  ],
  "implementation_hints": "To add a new auth check, follow the pattern in src/middleware/auth.ts line 45",
  "tokens_processed": 85000,
  "sub_calls_made": 8
}

Verification Checklist

Before returning FINAL():

• All relevant files identified
• Patterns extracted and documented
• Implementation hints are actionable
• Token count tracked
• Sub-call count within budget (≤20)
• Answer directly addresses the original query