Claude Code Workflows
Structured development with Claude Code. Think before coding, review before shipping.
Why
Claude Code is powerful but undisciplined. Left to its defaults, it:
- Jumps straight to writing code before researching the problem, discussing a solution, writing a plan.
- Might not review the code, or just review some aspects of it, leaving the codebase inconsistent with security or achitectural gaps.
- Loses context over long sessions — early decisions get forgotten, hallucinations increase, original goals drift
- Doesn't naturally follow follow a natural development flow: think → plan → build → review → ship → reiterate
- Produces limited or inconsisten auditable record of what was decided or implemented and why
Goal
The objective is to make Claude Code behave more like a disciplined senior engineer
Inspired by cc-sessions (DAIC workflow enforcement with trigger-phrase automation and sub-agents for Claude Code)
This project builds on top an opinionated workflow enforcement system with:
- Guardrails - Enforced workflow phases with gated steps.
- Guidance Structure to prompt and guide claude on each phase of development (coaching + skills)
- Accountability through plans, specs, review pipelines and traceability (github issues and observations)
How
Four tools that work together:
- Workflow Manager — Phase-based workflow enforcement with coaching and sequential gates
- Skills and Agents — skills for brainstorming, Agents for TDD, planning, debugging, code review
- claude-mem — cross-session memory via MCP server
- Status Line — context usage, git branch, workflow phase at a glance
Workflow
Six phases. Code edits are blocked until you discuss and approve a plan.
| Phase | Edits | What happens |
|---|
| OFF | Allowed | No enforcement |
| DEFINE | Blocked | Frame the problem, define outcomes |
| DISCUSS | Blocked | Research approaches, write plan |
| IMPLEMENT | Allowed | Execute plan with TDD |
| REVIEW | Allowed | Parallel review agents + verification |
| COMPLETE | Blocked | Validate outcomes, docs, open issues, handover |
Commands: /define /discuss /implement /review /complete /off /proposals /debug /autonomy
Any command can jump to any phase. /off closes the workflow. Soft gates warn when skipping steps.
See Architecture — Phase Model for the full breakdown: steps, agents, gates, milestones, and coaching per phase.
Autonomy Levels
Orthogonal to phase, the autonomy level controls how independently Claude operates. Set with /autonomy off|ask|auto (default: ask):
▶ Supervised (off): Step-by-step pair programming. Claude executes one plan step at a time, presents the change, and waits for your review before proceeding. Writes follow phase rules.▶▶ Semi-Auto (ask): Claude works freely within each phase but stops at phase boundaries for review and guidance before transitioning. No auto-commits.▶▶▶ Unattended (auto): Full autonomy. Claude auto-transitions between phases, auto-fixes review findings, auto-commits. Stops only when user input is genuinely needed or before git push.
Autonomy symbols are shown in the status line alongside the current phase.
Each cycle produces a plan (problem, approaches, rationale) and a spec (requirements, tasks, acceptance criteria).
Features
Docs
Installation
As a Claude Code Plugin (recommended)
Add the marketplace and install:
/plugin marketplace add azevedo-home-lab/claude-code-workflows
/plugin install workflow-manager
The plugin auto-wires hooks, installs the statusline, and initializes project state. No manual configuration needed.
Optional Tools
Installed separately with --iterm or --yubikey flags:
