English | 简体中文 | 日本語 | Español | 한국어 | Português (Brasil)

Claude Code can explore a codebase deeply. On non-trivial work, the harder problem is convergence. While designing an account-recovery flow, Claude may find a real inconsistency in token handling and spend most of the design on it, leaving the requested recovery behavior vague.

claude-code-workflows keeps that exploration pointed at an agreed result. It agrees on the outcome and exclusions before design, checks designs against the repository, verifies each task before commit, and, on larger changes, independently reviews the finished implementation to make sure it delivers the agreed result, does not include unnecessary changes, and has no serious functional, reliability, or security problems. Within that scope, Claude chooses the implementation details from the codebase.

Use Claude Code directly when the outcome and safe implementation boundary are already clear. Use these workflows when a change needs scope agreement, durable design decisions, a reliable handoff between contexts, or independent verification.

The workflow adds agent calls and artifacts, so it should earn that cost. Use it when a real side finding could pull a larger change away from its intended result, a design could be internally consistent but miss the requested behavior, or a passing test could fail to observe what it claims to prove.

Once the implementation scope is approved, Claude carries the tasks through focused verification, repository quality checks, commits, and final review without asking for routine implementation decisions. It asks the user only when the agreed product outcome or exclusions must change; Claude handles technical design and implementation choices. Because the process is packaged as a Claude Code plugin, a team can apply the same controls across repositories without prescribing Claude's steps.

Requires a Claude Code release with plugin marketplace support.

# 1. Start Claude Code
claude

# 2. Add the marketplace
/plugin marketplace add shinpr/claude-code-workflows

Install the plugin that matches your project. If the install tells you to run /reload-plugins, do that before invoking the recipe.

# Backend or general
/plugin install dev-workflows@claude-code-workflows
/recipe-implement "Add rate limiting to the public API"

# Frontend
/plugin install dev-workflows-frontend@claude-code-workflows
/recipe-front-design "Add account recovery screens"

# Full-stack
/plugin install dev-workflows-fullstack@claude-code-workflows
/recipe-fullstack-implement "Add user authentication with JWT + login form"

Install only one workflow plugin. dev-workflows-fullstack already contains the backend and frontend workflows. If you previously used full-stack recipes from dev-workflows, migrate to dev-workflows-fullstack.

/recipe-front-design stops after the applicable UI Spec and Design Doc are reviewed and approved. Run /recipe-front-plan and /recipe-front-build when you are ready to continue. For a backend or general change, /recipe-design, /recipe-plan, and /recipe-build provide the same staged path.

Claude Code supports project-scoped marketplaces and plugins. Commit the resulting .claude/settings.json so contributors are prompted to use the same workflow plugin.

claude plugin marketplace add shinpr/claude-code-workflows --scope project
claude plugin install dev-workflows-fullstack@claude-code-workflows --scope project

Replace dev-workflows-fullstack with the plugin that matches the repository. See the Claude Code plugin documentation for project and managed installation options.

flowchart LR
    A[Request] --> B[Agree on outcome and exclusions]
    B --> C{One evident implementation path?}
    C -->|Yes| S[Direct task cycle]
    S --> J[Complete]
    C -->|No| D[Inspect, design, and review]
    D --> E[Approve implementation scope]
    E --> F[Per task: implement, verify, quality-check, commit]
    F --> I[Independent implementation and security review]
    I -->|Correction| F
    I -->|Boundary changed| B
    I -->|Passed| J[Complete]

The number of product and design decisions determines the route, not file count or the amount of implementation work:

UI Specs, ADRs, and integration or E2E test skeletons appear only when their decisions or proof boundaries apply.

Generating an artifact does not advance the workflow on its own. Decision-changing design premises are resolved with observable evidence before approval, using a bounded probe only when it is the smallest sufficient proof.

The Work Plan is reviewed for coverage, dependency order, and executable verification before it authorizes implementation. Each task is committed only after its focused checks and applicable repository checks complete. When staged implementation is finished, separate reviews check the whole change against the agreed outcome, look for unnecessary changes and serious functional or reliability problems, confirm observable coverage, and assess security.

The main session decides which findings belong to the current outcome, resolves implementation questions from the repository, and keeps unaffected work moving. Review suggestions do not become work automatically. No change, removal, and reuse of what already exists are weighed before a fix that keeps or adds a mechanism, and that fix has to name the result the smaller answers cannot deliver. A correction can therefore remove something an earlier Design Doc or ADR selected, and the document is updated with it. An accepted correction returns through implementation and the affected verification gates.

Fresh contexts keep one phase's reasoning from silently becoming the next phase's authority. In the included Work Plan template each task cites the Design Doc, ADR, or UI Spec sections and acceptance criteria that constrain it, and the plan is finished only when every Design Doc obligation the implementation needs is covered by at least one task. It does not turn every document section or review suggestion into a task. An obligation left uncovered is a planning mistake, so the plan gains or adjusts a task instead of returning the question to the user.

- [ ] **P1-T1: Preserve the error response shape in the new handler**
  - **Source**: docs/design/example.md — API contract; AC-03
  - **Scope**: request handling for the affected endpoint
  - **Depends on**: none
  - **Executor lane**: backend
  - **Rollback boundary**: the handler change reverts with this task
  - **Verification**: existing contract test for the error path

The Task template carries binding decisions and observable contract values into implementation, each with a yes-or-no compliance check. After execution, the applicable repository checks run against the complete task change before commit. The final reviewers read the same approved sources and the completed code instead of relying on the implementation conversation. /recipe-quality-profile can record repository-specific quality rules and their sources in docs/project-context/quality.yaml; implementation executors and final reviewers use a confirmed profile alongside the approved sources.

The incremental sync feature in mcp-local-rag was a 42-file change across filesystem scanning, storage, and both the CLI and MCP surfaces. An independent security review sent the implementation back twice. It caught file reads happening before validation and a path-containment escape through a symlinked parent.

The run began with an existing Work Plan that referred to an ADR and Design Doc that were not present, leaving the approved source for its technical decisions unclear. The user chose to treat the Work Plan as the source of truth, and the recipe divided it into 13 planned tasks. The final implementation included the changes needed to verify the approved behavior, while the PR records why watch mode and persistent jobs were left out.

After the first run, inspect the artifacts:

• Did the agreed approach extend what already exists and give evidence for each addition?
• Can you follow each requirement into a task and an observable verification method?
• Did every completed task pass its focused and repository quality checks before commit?
• Did final review confirm that the whole change delivers the agreed outcome without unnecessary changes or serious functional, reliability, or security problems?
• When a reviewer proposed more work, did the report show why it was applied or declined?

/recipe-implement "Add rate limiting to the public API"

The recipe scopes the change, inspects the current implementation, creates only the documents required by its decisions, pauses when a decision is needed, and carries the plan through implementation and final review.

# Backend or general
/recipe-design "Design rate limiting for the public API"
/recipe-plan
/recipe-build

# React frontend
/recipe-front-design "Build a user profile dashboard"
/recipe-front-plan
/recipe-front-build

The design recipes inspect the existing code, confirm the scope, create the required documents, run an independent consistency review, and stop for approval. Planning and implementation can continue later, in a new context or by another contributor, from those approved artifacts.

The frontend path adds UI analysis and a UI Spec when UI structure or behavior remains to be designed, plus component architecture, React Testing Library, and TypeScript checks.

For example, two dashboard components may each handle loading correctly while the combined screen has no defined behavior when one is loading and the other has failed. The UI Spec records that state combination and traces it into design and test work before integration.

/recipe-fullstack-implement "Add user authentication with JWT + React login form"

When the change has multiple independent product outcomes, one PRD covers the whole feature. Backend and frontend design stay separate, design-sync checks the boundary between them, and the work plan uses vertical slices so integration is exercised before the end.

Use /recipe-fullstack-build to continue from an existing full-stack work plan. The full-stack plugin also includes the applicable backend and frontend recipes.

/recipe-review

The review workflow checks the completed implementation against the agreed outcome and repository standards, then runs an independent security review. Accepted corrections return to the appropriate implementation or document owner and are reviewed again.

/recipe-diagnose "API returns 500 on user login"

The diagnosis workflow maps execution paths, verifies suspected failure points, and presents solution trade-offs. It does not change the code.

/recipe-reverse-engineer "src/auth module"

This derives PRDs and Design Docs from the code and verifies the documents against the implementation. Use the full-stack option when the feature crosses backend and frontend.

For a walkthrough, see How I Made Legacy Code AI-Friendly with Auto-Generated Docs.

/recipe-front-adjust "Align the card spacing and actions with the design source"

The frontend plugin records how to reach the external design source, confirms the write set, and repeats visual verification until the adjustment passes its checks.

All workflow entry points use the recipe- prefix. Type /recipe- and use tab completion to see what the installed plugin provides.

The frontend plugin adds React-specific analysis, component architecture, React Testing Library, TypeScript checks, and applicable UI Spec generation from optional prototype code.

Specialized agents keep analysis and design separate from execution and final review. Each plugin includes only the roles its workflows use; the full-stack plugin combines the backend and frontend roles. The complete role list is folded below.

These agents are shared by the backend, frontend, and full-stack workflow plugins:

• Coding Principles. Code quality standards.
• Testing Principles. TDD, coverage, test patterns.
• Implementation Approach. Design decisions and trade-offs.
• Documentation Standards. Clear, maintainable docs.
• External Resource Context. Records how to reach design sources, design systems, API schemas, infrastructure definitions, and other resources outside the repository.
• LLM-Friendly Context. Clear prompts, handoffs, generated artifacts, and instructions for downstream agents.

Agents load these skills when the work calls for them. The frontend plugin also includes React and TypeScript-specific rules.

If you already have orchestration through custom prompts or CI and want only the best-practice guides, use dev-skills. If you want Claude to plan, execute, and verify a change end to end, install one of the workflow plugins instead.

• Minimal context footprint with no agents or recipe skills
• Coding, testing, design, and documentation guidance without a prescribed workflow
• Automatic skill loading when a task is relevant

Do not install dev-skills alongside a workflow plugin. They share the same skills, and duplicate descriptions can cause Claude Code to ignore skills after reaching its context limit.

/plugin install dev-skills@claude-code-workflows

To switch between plugin types:

# dev-skills -> dev-workflows
/plugin uninstall dev-skills@claude-code-workflows
/plugin install dev-workflows@claude-code-workflows

# dev-workflows -> dev-skills
/plugin uninstall dev-workflows@claude-code-workflows
/plugin install dev-skills@claude-code-workflows

These plugins cover adjacent work without changing the core development workflow:

• claude-code-discover: turns feature ideas into evidence-backed PRDs.
• metronome: detects shortcut-taking behavior and asks Claude to follow the defined procedure.
• linear-prism: validates requirements and turns them into structured Linear tasks.
• pr-review: reviews GitHub PRs against repository-specific criteria before posting approved findings.

/plugin install discover@claude-code-workflows
/plugin install metronome@claude-code-workflows
/plugin install linear-prism@claude-code-workflows
/plugin install pr-review@claude-code-workflows

Q: What if there are errors?

A: The quality-fixer agents handle test, type, lint, and build failures within the approved outcome, including adjacent changes required by the same responsibility or contract.

The workflow asks the user only when keeping both the requested outcome and its exclusions is no longer possible, or when an irreversible external action needs approval. It handles technical design, contracts, UI, architecture, persistence, and implementation changes on its own as long as they do not change what the product delivers.

Q: Is there a version for OpenAI Codex CLI?

A: Yes. codex-workflows provides the same workflow model, adapted to the Codex CLI environment.

Q: Should I commit the work plan and task files in docs/plans/?

A: No. Recipes treat docs/plans/ as ephemeral working state. Consumed task files and intermediate fix files are cleaned up after successful execution. The work plan may remain for review or a later build and can be deleted when it is no longer needed. Add the following line to your project's .gitignore so this working state stays out of git:

docs/plans/

PRDs, ADRs, UI Specs, and Design Docs live in their own directories (docs/prd/, docs/adr/, docs/ui-spec/, docs/design/) and are intended to be committed.

This marketplace supports the full lifecycle of building products with AI: product quality, discovery, implementation control, and verification. If your plugin helps developers build better products with AI coding agents, we'd like to hear from you.

See CONTRIBUTING.md for submission guidelines and acceptance criteria.

claude-code-workflows/
├── .claude-plugin/
│   └── marketplace.json        # Plugin definitions and per-plugin contents
├── agents/                     # Specialized analysis, design, execution, and review roles
├── skills/
│   ├── recipe-*/               # Workflow entry points
│   ├── documentation-criteria/ # Document rules and templates
│   ├── coding-principles/
│   ├── testing-principles/
│   ├── external-resource-context/
│   ├── llm-friendly-context/
│   └── ...
├── LICENSE
└── README.md

• Why LLMs Are Bad at 'First Try' and Great at Verification: why external feedback and fresh contexts are more reliable than asking one session to generate and judge its own work
• When Better Models Make Old Agent Workflows Worse: why the workflow is strict about boundaries and evidence without prescribing the route between them
• Reasoning Effort Is Not a Quality Setting: why broader exploration must still converge on the work the current outcome justifies
• Stop Putting Everything in AGENTS.md: why always-on instructions stay small while skills, design decisions, and task guidance are loaded where they apply

MIT License. Free to use, modify, and distribute.

See LICENSE for full details.

Built and maintained by @shinpr.