Bloat Detector Skill for Claude

Bloat Detector

Systematically detect and eliminate codebase bloat through progressive analysis tiers.

Bloat Categories

| Category | Examples |

|----------|----------|

| Code | Dead code, God classes, Lava flow, duplication |

| AI-Generated | Tab-completion bloat, vibe coding, hallucinated deps |

| Documentation | Redundancy, verbosity, stale content, slop |

| Dependencies | Unused imports, dependency bloat, phantom packages |

| Git History | Stale files, low-churn code, massive single commits |

Quick Start

Tier 1: Quick Scan (2-5 min, no tools)

/bloat-scan

Detects: Large files, stale code, old TODOs, commented blocks, basic duplication

Tier 2: Targeted Analysis (10-20 min, optional tools)

/bloat-scan --level 2 --focus code   # or docs, deps

Adds: Static analysis (Vulture/Knip), git churn hotspots, doc similarity

Tier 3: Deep Audit (30-60 min, full tooling)

/bloat-scan --level 3 --report audit.md

Adds: Cross-file redundancy, dependency graphs, readability metrics

When To Use

| Do | Don't |

|----|-------|

| Context usage > 30% | Active feature development |

| Quarterly maintenance | Time-sensitive bugs |

| Pre-release cleanup | Codebase < 1000 lines |

| Before major refactoring | Tools unavailable (Tier 2/3) |

When NOT To Use

• Active feature development
• Time-sensitive bugs
• Codebase < 1000 lines

Confidence Levels

| Level | Confidence | Action |

|-------|------------|--------|

| HIGH | 90-100% | Safe to remove |

| MEDIUM | 70-89% | Review first |

| LOW | 50-69% | Investigate |

Prioritization

Priority = (Token_Savings × 0.4) + (Maintenance × 0.3) + (Confidence × 0.2) + (Ease × 0.1)

Module Architecture

Tier 1 (always available):

• See modules/quick-scan.md - Heuristics, no tools
• See modules/git-history-analysis.md - Staleness, churn, vibe coding signatures
• See modules/growth-analysis.md - Growth velocity, forecasts, threshold alerts

Tier 2 (optional tools):

• See modules/code-bloat-patterns.md - Anti-patterns (God class, Lava flow)
• See modules/ai-generated-bloat.md - AI-specific patterns (Tab bloat, hallucinations)
• See modules/documentation-bloat.md - Redundancy, readability, slop detection
• See modules/static-analysis-integration.md - Vulture, Knip

Shared:

• See modules/remediation-types.md - DELETE, REFACTOR, CONSOLIDATE, ARCHIVE

Ecosystem-Level Detection

Patterns that span plugin boundaries or manifest configuration,

discovered through ecosystem-wide audits.

alwaysApply Accumulation

Flag plugins with 3+ skills where alwaysApply: true.

Each always-on skill injects its full text into every session,

creating a baseline token floor before the user types anything.

Sum the estimated_tokens fields to report total per-session cost.

Hook Registration Gaps

Compare hooks declared in plugin.json or openpackage.yml

against entries in hooks.json. A hook present in hooks.json

but absent from the manifest is invisible to the plugin loader

and cannot be audited, versioned, or disabled through normal

plugin management.

Boilerplate Footer Detection

Scan skill files for identical multi-line text blocks repeated

across 10+ files (e.g., generic troubleshooting sections like

"Command not found / Permission errors / Unexpected behavior").

These are copy-paste artifacts that inflate token cost without

adding skill-specific value.

ToC Bloat in Skills

Skills loaded into model context gain nothing from HTML-style

Tables of Contents. Detect ## Table of Contents followed by

bulleted anchor-link lists. These waste tokens since

the model reads sequentially, not via hyperlinks.

Unregistered Module Subdirectories

Compare files on disk in skills/*/modules/ against the

modules: list in each skill's SKILL.md frontmatter. Files

that exist on disk but are not listed in the manifest are

invisible to progressive loading and may be dead weight or

missing from the load path.

Auto-Exclusions

Always excludes: .venv, __pycache__, .git, node_modules, dist, build, vendor

Also respects: .gitignore, .bloat-ignore

Safety

• Never auto-delete - all changes require approval
• Dry-run support - --dry-run for previews
• Backup branches - created before bulk changes

Related

• bloat-auditor agent - Executes scans
• unbloat-remediator agent - Safe remediation
• context-optimization skill - MECW principles
• /bloat-scan command - User-facing slash command that invokes this

skill; there is no separate bloat-scan skill

Exit Criteria

• [ ] At least one tier of scan (Tier 1, 2, or 3) completed and

findings listed per bloat category (Code, AI-Generated,

Documentation, Dependencies, Git History)

• [ ] Each finding assigned a confidence level (HIGH/MEDIUM/LOW) and

a priority score computed from the formula

`(Token_Savings × 0.4) + (Maintenance × 0.3) + (Confidence × 0.2)

+ (Ease × 0.1)`

• [ ] No auto-deletions executed without explicit user approval;

every proposed removal documented as DELETE, REFACTOR,

CONSOLIDATE, or ARCHIVE per modules/remediation-types.md

• [ ] Ecosystem-level patterns checked (alwaysApply accumulation,

hook registration gaps, boilerplate footers, ToC bloat) if the

scan target contains plugin directories