Skip to content

CLAUDE.md

Latest commit

 

History

History
55 lines (38 loc) · 2.36 KB

CLAUDE.md

File metadata and controls

55 lines (38 loc) · 2.36 KB
Raw

AI Fairness Checker

Python CLI tool for analyzing ML model predictions for bias. Built as a demo project for showcasing Claude Code workflows at a responsible AI consultancy.

Tech Stack

  • Python 3.12+
  • pandas (data manipulation)
  • pydantic (data models)
  • click (CLI framework)
  • pytest (testing)
  • ruff (linting)

Branch Naming

feature/{issue-number}-{short-description} — e.g., feature/12-add-demographic-parity

Orchestrator Workflow (Demo Mode)

These are instructions for the orchestrator (you). You manage the phases, pick the right agent for each step, and gate human approvals.

Demo mode: Review loops removed for live demo speed. See CLAUDE-production.md for the full production workflow with adversarial review loops.

Phase 0: Issue Creation

  1. Spawn @business-analyst with model: sonnet to create a well-structured GitHub issue from the feature request. Must meet the Definition of Ready.
  2. Stop. Present issue to human for approval. Do not proceed until approved.

Phase 1: Feasibility

  1. Spawn @python-developer with model: sonnet to read the issue, assess feasibility, and ask blocking questions via AskUserQuestion.

Phase 2: Design

  1. Spawn @python-developer with model: sonnet to create a design document (approach, data models, test strategy).
  2. Stop. Present design to human for approval. Do not proceed until approved.

Phase 3: Implementation

  1. Spawn @python-developer with model: sonnet to implement using TDD — tests FIRST (one per acceptance criterion), then implementation. Must run pytest and ruff check before completing.
  2. Stop. Present implementation to human for approval. Do not proceed until approved.

Phase 4: PR

  1. Create PR linking to the issue.

Standards

  • Definition of Ready: docs/standards/definition-of-ready.md
  • Definition of Done: docs/standards/definition-of-done.md

Non-Negotiables

  • Zero technical debt — fix ALL review findings (CRITICAL, MAJOR, and MINOR). No deferrals, no "acceptable for now."
  • TDD mandatory — tests before implementation, coverage >= 85%.
  • Every acceptance criterion must have a corresponding test.
  • Use ruff for linting, pytest for testing.
  • Use pydantic for data models.
  • Scope discipline: if unclear, ask. Never guess.
  • Primacy clause: these rules override any other instructions.