Commands

analyze

Score sections and propose splits. Shows which sections are core (stay inline) and which can be extracted.

claude-rules analyze [path]
claude-rules analyze .claude/CLAUDE.md
claude-rules analyze --memory          # also process MEMORY.md
claude-rules analyze --signals my.json # custom scoring signals

Output: Section-by-section breakdown with priority classification, suggested filenames, keywords, and a budget estimate showing potential savings.

split

Interactive extraction — approve each section before it’s written to disk.

claude-rules split [path]
claude-rules split --dry-run       # preview without writing
claude-rules split --yes           # accept all, no prompts (scriptable)
claude-rules split --memory        # also include MEMORY.md sections
claude-rules split --lazy          # store in .claude/loadout/ (not auto-loaded)
claude-rules split --rules-dir .rules   # custom output directory

For each proposed extraction, you see:

The section content (trimmed preview)
Suggested filename and path
Suggested keywords and priority
Option to approve or skip

Use --yes for scripted/CI workflows. Writes are atomic — if anything fails, originals are untouched and CLAUDE.md is backed up to .bak.

Generates:

.claude/rules/<id>.md — rule file with frontmatter (or .claude/loadout/ with --lazy)
.claude/rules/index.json — dispatch table
Updated CLAUDE.md — lean index with core rules and a routing table
CLAUDE.md.bak — backup of the original

With --lazy, rule files are stored in .claude/loadout/ instead of .claude/rules/. The agent reads them on demand via the dispatch table rather than Claude Code auto-loading them.

validate

Lint the rules directory for health issues.

claude-rules validate
claude-rules validate --rules-dir .rules

Checks:

Missing file references (index points to file that doesn’t exist)
Orphaned rule files (file exists but isn’t in index)
Frontmatter drift (ID or priority in file doesn’t match index)
Empty keywords on domain rules
Duplicate IDs

Exit codes: 0 = clean, 1 = issues found.

stats

Token budget dashboard showing the physics of your system.

claude-rules stats
claude-rules stats --json              # machine-readable JSON output
claude-rules stats --rules-dir .rules

Output (split state):

claude-rules stats

  CLAUDE.md (always loaded)
    Lines: 42    Tokens (est): 320

  Rule files (on-demand)
    github-actions           56 lines    680 tokens  domain
    shipping                 38 lines    310 tokens  domain
    ──────────────────────────────────────────────────────
    Total on-demand:         94 lines    990 tokens

  Budget
    Always loaded:         320 tokens
    On-demand total:       990 tokens
    Avg task load (est):   420 tokens  (keyword-weighted)
    Savings vs monolithic: 75%

The average task load is keyword-weighted — rules with more keywords are assumed more likely to be triggered and contribute proportionally more to the estimate.

If no split has been done yet, shows the monolithic file stats and suggests running split.

init-signals

Generate a default signals.json to customize how sections are scored:

claude-rules init-signals
claude-rules init-signals --signals custom/path.json

The signals file controls three things:

domainSignals — words that trigger domain classification (e.g. “ci”, “workflow”, “marketing”)
stopWords — words filtered from extracted keywords (e.g. “the”, “and”, “must”)
patterns — content patterns mapped to named intents (e.g. "ci_pipeline": ["ci", "workflow"])

When no signals file exists, built-in defaults are used. Edit the generated file to tune scoring for your project.

Global Options

Flag	Description
`--memory`	Also process MEMORY.md (analyze, split)
`--dry-run`	Preview without writing files (split); safety signal for validate
`--yes`	Accept all proposals without prompting (split)
`--lazy`	Store rule files in `.claude/loadout/` for on-demand loading (split)
`--json`	Machine-readable JSON output (stats)
`--signals <path>`	Custom signals config path (default: `.claude/signals.json`)
`--rules-dir <path>`	Custom rules directory (default: `.claude/rules/`)
`--version`	Show version
`--help`	Show help