tech-debt-detector

name	tech-debt-detector
description	Detect and classify technical debt in AI-generated code — patterns specific to LLM outputs, shallow implementations, missing edge cases, and accumulation signals.
version	1.0.0
last-updated	2026-04-22
model_tested	claude-sonnet-4-6
category	eval
platforms	["claude-code","codex","gemini-cli","cursor","copilot","windsurf","cline"]
language	en
geo_relevance	["global"]
priority	high
dependencies	{"mcp":[],"skills":[],"apis":[],"data":[]}
update_sources	[{"url":"https://newsletter.pragmaticengineer.com/p/the-impact-of-ai-on-software-engineers-2026","check_frequency":"quarterly","last_checked":"2026-04-22"}]
license	MIT

Technical Debt Detector (AI-Generated Code)

Research shows a 2500% increase in code defects linked to AI-generated code (BCG 2026). This skill identifies debt patterns specific to LLM outputs.

When to Use

• After generating code with AI (review before merging)
• During code review of AI-assisted PRs
• When refactoring AI-generated modules
• When investigating production issues in AI-written code
• Periodic tech debt audits

AI-Specific Debt Patterns

Pattern 1: Shallow Implementation

LLMs produce code that works for the happy path but fails on edges.

Signals:

• No error handling beyond generic try/catch
• No input validation at system boundaries
• No null/undefined checks on external data
• No timeout on network calls
• Functions that work for sample data but fail at scale

Check: For each function, ask: "What happens with empty input? Null? Very large input? Concurrent access? Network failure?"

Pattern 2: Over-Abstraction

LLMs love creating abstractions even when unnecessary.

Signals:

• Wrapper classes with no added logic
• Factory patterns for single implementations
• Interface with exactly one implementation
• Helper functions called only once
• Generic framework for a specific problem

Check: "Can I delete this abstraction and use the concrete implementation directly?"

Pattern 3: Stale Patterns

LLMs use patterns from training data that may be outdated.

Signals:

• Class components in React (should be hooks)
• Callbacks instead of async/await
• var instead of const/let
• jQuery patterns in modern codebase
• Deprecated API usage

Check: "Is this the current recommended pattern for this framework version?"

Pattern 4: Copy-Paste Drift

LLMs generate similar code for similar tasks without deduplication.

Signals:

• 3+ functions with >70% similar logic
• Same validation logic repeated in multiple places
• Similar error messages with slight variations
• Duplicate type definitions

Check: "Are there 3+ places doing essentially the same thing?"

Pattern 5: Missing Observability

LLMs rarely add logging, metrics, or monitoring.

Signals:

• No logging in error paths
• No metrics on key operations
• No health check endpoints
• No structured error codes
• No request tracing

Check: "If this fails in production at 3 AM, can I diagnose the problem from logs alone?"

Pattern 6: Hardcoded Configuration

LLMs often hardcode values that should be configurable.

Signals:

• URLs, ports, timeouts in code (not config)
• Magic numbers without named constants
• Environment-specific values in source
• Feature flags as if/else in code

Check: "Can I deploy this to a different environment without changing code?"

Debt Classification

Severity	Impact	Fix Timeline
Critical	Security vulnerability, data loss risk	Immediately
High	Production failures under load or edge cases	This sprint
Medium	Maintainability issues, duplication	Next sprint
Low	Style, naming, minor abstractions	Backlog

Output Format

TECH DEBT SCAN: {file/module}
Generated by: {AI tool if known}
Patterns found: {count}

[HIGH] Pattern 1 (Shallow): No error handling in fetchUserData() (line 42)
[MED]  Pattern 4 (Copy-Paste): validateInput() duplicated in 3 controllers
[LOW]  Pattern 3 (Stale): Using deprecated fetch API options (line 78)

Estimated debt: {hours to fix}
Priority: Fix HIGH items before merging.

What This Skill Does NOT Do

• Does not fix the debt (identifies and classifies)
• Does not replace SonarQube or ESLint (complements with AI-specific patterns)
• Does not measure test coverage (use coverage tools)
• Does not block PRs (advisory only)