claude-workflow

Disciplined agent workflow system for Claude Code. Linear as source of truth, Gherkin specs, TDD execution, deterministic verification.

Philosophy

Writing code is cheap. Trust, correctness, and maintainability are expensive.

• Agents start fresh each session (no persistent mental model)
• Specs drift from implementation
• No rationale trail for decisions
• Hard to verify agent output is correct

Solution: Linear is source of truth, Gherkin specs are executable, verification is deterministic, humans review decisions—not implementation.

Deterministic Verification

Trust tool output, not LLM claims.

Issue Templates

Work Unit (Sub-issue)

Scenario: [Clear behavior description] Given [precondition] When [action] Then [expected result]

Multi-repo features: Create separate Work Units for each repo with dependencies. Example: BHU-100: API endpoint (backend) → BHU-101: Mobile integration (mobile, blocked by BHU-100)