entropy-reduction

name	entropy-reduction
description	USE THIS SKILL whenever the user wants to: refactor a codebase, clean up tech debt, improve code consistency, reduce complexity, make a codebase more maintainable, or when the user mentions entropy, code quality, code health, or asks to clean up or simplify a module, directory, or entire repo. Especially valuable for AI-first and vibe coding teams where code accumulates disorder quickly.

Systematically reduce codebase disorder without breaking existing behavior. Identify high-entropy areas and lower entropy through safe, incremental refactoring. Prioritize obviously high-entropy areas over minor style issues.

Entropy Axes

Axis	What it means	Typical symptoms
Structural	Layering, dependencies, topology	Oversized files/functions, util black holes, circular deps, layer violations
Semantic	Naming, concepts, state representation	Same concept with multiple names, scattered state machines, ad-hoc data models
Behavioral	API contracts, error handling, boundary behavior	Inconsistent error responses, different status codes for the same error class, env-dependent behavior
Evolutionary	Change patterns, ownership, understandability	Giant mixed-purpose PRs, no clear module ownership, design intent buried in chat history

Workflow

When given a codebase scope (file, directory, module, or repo), follow these four phases:

Phase 1: Diagnose

Scan the target scope and identify high-entropy symptoms. For each finding, note:

• What: the specific symptom (e.g., "5 different names for the user concept across 3 modules")
• Which axis: structural / semantic / behavioral / evolutionary
• Severity: how much disorder this creates for future maintainers

Concrete things to look for:

Structural entropy

• Files over ~500 lines or functions over ~50 lines
• util/, common/, helpers/ directories that have become catch-all dumping grounds with business logic mixed in
• Circular dependencies between modules (A→B→C→A)
• UI code directly importing from infrastructure/database layers, skipping service layers
• God objects or god modules that everything depends on

Semantic entropy

• The same domain concept referred to by multiple names (e.g., user / member / account / customer / profile all meaning the same thing)
• The same business state represented differently in different places (enum in one module, string in another, integer in a third)
• State transition logic scattered across handlers, services, frontend, and batch jobs instead of living in one place
• Ad-hoc structs/dicts/maps used where a proper domain model would clarify intent

Behavioral entropy

• The same type of error returning different HTTP status codes or payload shapes in different endpoints
• Boundary cases (empty data, missing params, permission denied) handled inconsistently across similar APIs
• Behavior that diverges significantly between local/test/production environments

Evolutionary entropy

• Recent PRs that mix feature work, refactoring, and formatting in a single changeset
• Modules with no clear ownership — multiple authors, no consistent patterns
• Design decisions whose rationale exists only in chat logs or memories, not in code or docs

Present findings as a concise list — developer-readable, no jargon overload.

Phase 2: Plan

Design an incremental entropy reduction plan:

1. Propose 1–3 small, safe changes that can be done immediately. For each:

   • State the goal (what entropy it reduces)
   • State the risk (what could break)
   • Label the entropy axes it addresses
   • Estimate scope: is this local (single file/module) or cross-cutting?

2. Prioritize by: highest entropy reduction per unit of risk. Do the safe, high-impact changes first.

3. Flag anything that needs human input — if you're unsure about a naming convention, an architectural boundary, or whether a behavior change is acceptable, ask rather than guess.

Phase 3: Refactor

Execute the planned changes:

• Preserve external behavior. Every refactor must be behavior-preserving unless explicitly approved otherwise.
• Work in small steps. Each change should be independently committable and reversible.
• When renaming or moving code, ensure all references are updated. Check imports, type references, tests, and config files.
• For risky changes (anything that touches shared interfaces or public APIs):
  • List the callers/consumers that could be affected
  • Suggest specific test cases to verify the change is safe
• In commit messages, briefly explain the entropy reduction intent (e.g., "Consolidate user/member/account naming to 'user' — reduces semantic entropy across auth and billing modules").

Phase 4: Verify

After making changes:

• Run existing tests if available. All must pass.
• If no tests exist for the changed code, note which test cases should be added to protect the refactored behavior.
• Confirm that no new cross-layer dependencies, naming variants, or util dumping has been introduced.

Constraints

These are hard rules — do not violate them:

1. No large rewrites of stable code. If it works and isn't causing maintenance pain, leave it alone.
2. No new util/common black holes. Do not create new catch-all modules. Move business logic into the domain it belongs to.
3. No new naming variants. If the codebase calls it user, don't introduce member or account as an alternative name for the same concept.
4. No new cross-layer dependencies. Respect the dependency direction: UI → Service → Domain → Repository → Infrastructure.
5. Reuse existing abstractions. Before creating a new helper, check if one already exists. Three similar lines of code is better than a premature abstraction.
6. When uncertain, ask. Do not force an abstraction or rename you're unsure about. State your assumption and ask the human to confirm.

Output Format

Structure your output as:

1. Diagnosis — Brief list of high-entropy findings with axis labels and severity
2. Plan — Numbered list of proposed changes with goals, risks, and axis labels
3. Changes — The actual code modifications (complete, compilable/runnable)
4. Test suggestions — Any tests that should be added to protect refactored behavior
5. Follow-up — Larger entropy reduction opportunities that need more planning or human input

Keep the diagnosis and plan sections concise. The code changes are the main deliverable.