Wizard

Wizard - 8-Phase Implementation

Goal

Deliver a production-quality implementation with zero regressions. Success = all new tests pass, full test suite passes, linter clean, and user has reviewed and approved a complete diff summary.

Dependencies

• Tools: Read, Write, Edit, Bash, Grep, Glob
• Requires a test runner (language-appropriate: pytest, jest, go test, etc.)
• Reads CLAUDE.md for project conventions at Phase 1

Context

Reads CLAUDE.md and any linked specs or issues at Phase 1. References existing codebase patterns discovered during Phase 2 exploration.

A production-tested methodology that prevents common AI coding failures.

Phase 1: Plan

• Read CLAUDE.md and any linked specs/issues
• Assess complexity: files affected, architectural impact, risk level
• Build structured task list before writing any code
• Present plan for user approval

Phase 2: Explore

• Grep for every model, method, class, and constant you intend to use
• Verify they actually exist before referencing them
• Read related code to understand patterns and conventions
• This prevents hallucinated method chains and non-existent APIs

Phase 3: Test First

• Write failing tests covering:
  • Happy path
  • Edge cases (null, empty, boundary values)
  • Error conditions
• Use mutation-resistant assertions (specific values, not truthiness)
• Run tests - they MUST fail
• Note the failing tests (don't modify them later to make them pass)

Phase 4: Implement Minimum

• Write only the code needed to make tests pass
• No extra features, no premature abstractions
• No "while I'm here" improvements
• Keep the implementation as simple as possible

Phase 5: Verify No Regression

• Run the broader test suite, not just new tests
• Catch unrelated breakage immediately
• Fix any regressions before moving forward

Phase 6: Document

• Add inline comments only where logic isn't self-evident
• Update any relevant documentation files
• Context is freshest now - capture what future readers need

Phase 7: Adversarial Review

Think like an attacker, not the author:

• What happens under concurrent execution?
• What if inputs are null, empty, or malformed?
• Are there hardcoded values that should be configurable?
• Does this handle production scale?
• Are there security implications?

Fix anything found before proceeding.

Phase 8: Quality Gate

• Run full test suite one final time
• Run linter if configured
• Review the complete diff
• Present summary of all changes to user

Output

• No dedicated output file. Deliverable is working code committed to the project.
• Phase 8 produces a diff summary presented to the user for final approval before closing.
• CHECKPOINT (Phase 1): Present the task plan to the user. Do NOT write code until they approve.
• CHECKPOINT (Phase 8): Present the complete diff summary. Confirm user is satisfied before marking done.

When to Skip Phases

• Quick typo fix: Skip to Phase 4
• Pure refactor with existing tests: Skip Phase 3
• Exploratory/prototype: Skip Phases 6-8

State which phases you're skipping and why.