Skip to content
cccskills
FreePublished on GitHub

nw-refactor

Applies the Refactoring Priority Premise (RPP) levels L1-L6 for systematic code refactoring. Use when improving code quality through structured refactoring passes.

The English summary is being prepared. Showing the original description.

See how to install

Files included (1)

  • SKILL.md4.7 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

NW-REFACTOR: Systematic Code Refactoring

Wave: CROSS_WAVE Agent: Crafty (nw-software-crafter) Command: *refactor

Overview

Applies the Refactoring Priority Premise (RPP) — 6-level hierarchy L1 Readability|L2 Complexity|L3 Responsibilities|L4 Abstractions|L5 Design Patterns|L6 SOLID++. For complex multi-class refactorings, agent applies Mikado Method internally.

Execution Model — Batch-Then-Verify (default, unconditional)

/nw-refactor runs batch-then-verify by default, regardless of test-suite speed:

  1. Cascade governs PLANNING ORDER only — analyze and plan lower levels before higher (L2 transformations build on L1, L3 on L2, etc.). The cascade is a planning discipline, not a test-gating sequence.
  2. Edits applied as one coherent batch — apply all planned L1-L6 transformations in a single editing session. NO interleaved test runs between levels.
  3. Run the test suite exactly ONCE, at the very end, after all L1-L6 edits.
  4. If RED after the batch: diagnose the breakage and fix the production code. Do NOT modify tests to make them pass. A test that must change to pass means either (a) the refactor altered observable behavior — revert that transformation — or (b) the test encoded an implementation detail rather than behavior — flag this to the user explicitly before touching it. Tests changing during a refactor is a signal, not a step.

Legacy incremental variant (opt-in only): the nw-progressive-refactoring skill documents the incremental L1→test→L2→test cycle. It is NOT the default — use it only when explicitly requested. Anchor: feedback_refactor_batch_when_test_suite_slow_2026_05_19 (the prior "only batch when suite slow" conditional is removed — batch is now unconditional).

Context Files Required

  • src/* - Production codebase
  • tests/* - Test codebase

Agent Invocation

@nw-software-crafter

Execute *refactor for {target-class-or-module}.

Context Files:

  • src/*
  • tests/*

Configuration:

  • level: 3 # Shorthand: --from=1 --to=3 (RPP range)
  • from: 1 # Start RPP level (default: 1)
  • to: 3 # End RPP level (default: same as level)
  • scope: module # file/module/project
  • method: extract # extract/inline/rename/move
  • mikado_planning: false # Use Mikado Method for complex refactorings

Progress Tracking

The invoked agent MUST create a task list from its workflow phases at the start of execution using TaskCreate. Each phase becomes a task with the gate condition as completion criterion. Mark tasks in_progress when starting each phase and completed when the gate passes. This gives the user real-time visibility into progress.

Success Criteria

  • Code quality metrics improved (measured before/after)
  • L1-L6 transformations planned in cascade order, applied as one batch
  • Test suite run exactly once, at the end — GREEN
  • If RED after batch: production code fixed (tests NOT modified to pass)
  • Technical debt reduced measurably

Next Wave

Handoff To: {invoking-agent-returns-to-workflow} Deliverables: Refactored codebase with quality improvements

Examples

Example 1: Module-level readability refactor

/nw-refactor src/auth/token_manager.py --level=2 --scope=module

Crafty applies RPP L1-L2: rename ambiguous variables|extract magic numbers into constants|remove dead code (L1), then simplify conditionals|extract long methods (L2).

Example 2: SOLID-level design refactor

/nw-refactor src/billing/ --level=6 --scope=module --mikado_planning=true

Crafty uses Mikado Method for multi-class refactoring, applies dependency inversion|interface segregation across billing module.

Example 3: RPP range sweep (L1-L3)

/nw-refactor src/des/domain/ --from=1 --to=3 --scope=module

Sweeps L1 readability|L2 complexity|L3 responsibility smells. Cascade governs planning order (analyze L1 before L2 before L3); edits applied as one batch; suite run once at the end.

Example 4: Targeted single-level refactor

/nw-refactor src/des/cli/verify.py --level=3 --scope=file

Targets L3 responsibility smells only (Large Class, Feature Envy, Shotgun Surgery). Assumes L1-L2 already clean.

Expected Outputs

src/*                              (refactored production code)
tests/*                            (refactored test code)
docs/refactoring/
  refactoring-log.md
  quality-metrics.md

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

Review dimensions for validating agent quality - template compliance, safety, testing, and priority validation

The English summary is being prepared. Showing the original description.

nWave-ai/nWave620Updated Sep 16, 2026

Review dimensions for validating agent quality - template compliance, safety, testing, and priority validation

The English summary is being prepared. Showing the original description.

nWave-ai/nWave620Updated Sep 16, 2026

Review dimensions for acceptance test quality - happy path bias, GWT compliance, business language purity, coverage completeness, walking skeleton user-centricity, priority validation, observable behavior assertions, traceability coverage, and walking skeleton boundary proof

The English summary is being prepared. Showing the original description.

nWave-ai/nWave620Updated Sep 16, 2026

Detailed 5-phase workflow for creating agents - from requirements analysis through validation and iterative refinement

The English summary is being prepared. Showing the original description.

nWave-ai/nWave620Updated Sep 16, 2026

5-layer testing approach for agent validation including adversarial testing, security validation, and prompt injection resistance

The English summary is being prepared. Showing the original description.

nWave-ai/nWave620Updated Sep 16, 2026

Architectural style selection decision matrices, trade-off analysis, structural enforcement rules, and combination patterns. Load when choosing or evaluating architecture styles.

The English summary is being prepared. Showing the original description.

nWave-ai/nWave620Updated Sep 16, 2026

See all skills by nWave-ai

Report a problem with this skill