Plan-Code-Review Workflow
One workflow. Four AI assistants. Production-ready code—without the guesswork.
Apply the compounding dev cycle (Plan → Code → Review/Test) to Cursor, Claude Code, GitHub Copilot, or Codex. Get a single source of truth, automatic code review, and a repeatable path from idea to shipped feature.
The compounding dev cycle (why this workflow)
AI can write code fast—but without a clear plan and review loop, you get scope creep, untested assumptions, and “what did it actually build?” moments. The compounding dev cycle fixes that.
| Phase |
What happens |
What you get |
| Plan |
One written plan: scope, acceptance criteria, technical approach |
No guesswork. Code implements to the plan; review verifies against the same criteria. |
| Code |
Implementation + tests + implementation notes |
Traceability. Every change links back to the plan (e.g. “implements AC-1”). |
| Review/Test |
Automatic code review (backend + frontend reviewers) |
Quality gates. Critical issues → rework plan → Code → Review again until production ready. |
| Plan (next) |
Rework or new scope becomes the next cycle |
Clear handoffs. Written artifacts for every phase so you and the AI always have the right context. |
In practice: Run /feature-plan once → get a plan file. Run /project-manager with that plan → implementation runs, then review runs automatically. If reviewers find critical issues, the workflow re-plans, fixes, and re-reviews until the code is ready. One plan, one command, production ready.
Modes in each assistant: Plan, Agent, and Ask
Cursor, Claude Code, and GitHub Copilot (VS Code) each provide Plan-, Agent-, and Ask-style modes. You can align the compounding cycle with these native modes:
| Cycle phase |
Cursor |
Claude Code (VS Code) |
GitHub Copilot (VS Code) |
| Plan |
Plan mode — plan only, no edits |
Plan mode — describes what it will do, waits for approval before changes |
Plan agent — structured plan (Discovery → Alignment → Design → Refinement), then hand off to Agent |
| Code |
Agent mode — full implementation |
Normal or Auto-accept — implements with or without per-step approval |
Agent — autonomously plans and implements across files, runs commands |
| Review/Test |
Ask mode — read-only feedback, rework list |
Plan mode or Normal (reject edits) — use for review-only; approve only when applying fixes |
Ask — answers and explains without making file changes; use for review feedback |
How to use: For Plan, use each assistant’s Plan mode/agent so it only produces a plan (or describes steps and waits). For Code, use Agent (or Normal/Auto-accept in Claude). For Review/Test, use Ask (or Plan/Normal in Claude with “produce rework list, do not apply”) so you get a rework list without unwanted edits. Then loop: rework plan → Code again → Review again.
Codex: The Codex IDE/CLI also supports plan-before-execute and agent behavior; the extension wires the cycle via AGENTS.md and skills. Check Codex documentation for current mode names and shortcuts.
Docs: Cursor modes · Claude Code permission modes (Plan, Normal, Auto-accept) · GitHub Copilot Chat modes (Ask, Edit, Agent) · VS Code Copilot agents (Ask, Plan, Agent).
Why install?
- No more “what did the AI build?” — One plan doc is the contract. Code and review both use it.
- Automatic code review — Backend and frontend reviewers run after implementation; critical issues trigger a rework loop until gates pass.
- Cursor mode discipline — Cursor, Claude Code, and GitHub Copilot each have Plan-, Agent-, and Ask-style modes. Use Plan for planning only, Agent for implementation, Ask for review-only feedback; the extension’s rules and instructions align the compounding cycle with each assistant’s native modes.
- Works with your stack — Same cycle for Cursor (slash commands), Claude Code, Copilot, or Codex; the extension installs the right files for your assistant.
- 17 agents + 19 skills — Planning, architecture, API design, security, accessibility, E2E—all wired into the cycle.
- Takes minutes — Install the extension, run one “Apply workflow” command, and you’re set.
Token-efficient outputs (new in 5.1.0)
The workflow now includes an always-on token-efficiency policy across Cursor, Claude Code, GitHub Copilot, and Codex workflow docs.
- Default behavior: concise but complete responses.
- Levels:
lite, full (default), ultra.
- Safety override: high-risk/destructive/security-sensitive instructions switch to explicit full-language warnings before returning to compact mode.
- Risk controls included: clarity regression, cross-assistant drift, output-format conflicts, and level inconsistency.
This keeps responses shorter without weakening safety, required structure, or technical accuracy.
Credit: Token-efficiency style is inspired by the Caveman skill by Julius Brussee: Caveman SKILL.md.
Install the extension
- VS Code: VS Code Marketplace or Open VSX — search for Plan-Code-Review Workflow.
- Cursor IDE: The Cursor plugin is not yet available in Cursor’s plugin marketplace (approval pending). Use this instead: install the extension in VS Code, open your project in VS Code, run Plan-Code-Review: Apply workflow for Cursor, then open the same project in Cursor—the workflow files (
.cursor/, .cursor-plugin/) are in your repo and Cursor will use them. Full plugin install from Cursor’s marketplace will be available once approved.
Apply the workflow to your project
- Open your project root in VS Code (or in Cursor if you already have the workflow applied).
- Command Palette (
Ctrl+Shift+P / Cmd+Shift+P) → run one of:
- Plan-Code-Review: Apply workflow for… (choose AI) — pick Cursor, Claude Code, Copilot, or Codex
- Or a specific command: Plan-Code-Review: Apply workflow for Cursor / Claude Code / GitHub Copilot / Codex
For Claude Code, you’ll be asked whether to install in the project root (.claude in this workspace) or in your user directory (~/.claude/, so the workflow applies to all projects).
Cursor users: Install the extension in VS Code, open your project there, run Plan-Code-Review: Apply workflow for Cursor, then switch to Cursor and open the same project—you get full slash commands and agents. No Cursor plugin install needed until the plugin is approved.
The extension writes the right files into your workspace. Your AI assistant uses them automatically.
Remove the workflow
To remove only what this extension added (your existing commands, skills, rules, and hooks from other sources stay untouched):
- Command Palette → Plan-Code-Review: Remove workflow from all AI assistants
The extension keeps a manifest of what it applied (workspace and user). Remove uses that list so only extension-added items are deleted. To see exactly what’s recorded before removing:
- Command Palette → Plan-Code-Review: Show applied workflow log
| Assistant |
What gets created |
| Cursor |
.cursor/, .cursor-plugin/ — full slash commands, agents, rules, hooks |
| Claude Code |
.claude/agents/, .claude/rules/, .claude/skills/, CLAUDE.md, .claude/hooks/ — in project root or in ~/.claude/ (you choose when applying) |
| GitHub Copilot |
User-level: rules, agents, skills in prompts folder (all workspaces). Workspace: .github/ — agents/ (custom agents as <name>.agent.md, skills inlined), copilot-instructions.md (rules), instructions/compounding-dev-cycle.instructions.md (path-specific, applyTo **), workflows/ (hooks), ISSUE_TEMPLATE/ (structured prompts); AGENTS.md at repo root (references cycle + custom agents). |
| Codex |
User-level: ~/.codex/AGENTS.md (compounding dev cycle) + ~/.codex/skills/ (synced). Workspace: AGENTS.md |
Codex: Codex has no separate rules folder; it uses only the AGENTS.md instruction chain (global ~/.codex/AGENTS.md and workspace AGENTS.md). To use an alternate filename (e.g. RULES.md), add it to project_doc_fallback_filenames in ~/.codex/config.toml (see Codex customization).
Cursor: Use /feature-plan, /project-manager, /api-new, etc. Others: Reference the generated files in chat (e.g. “Follow CLAUDE.md”, “Use the code-review skill”).
What’s inside
Commands (Cursor): /feature-plan, /project-manager, /new-task, /api-new, /api-test, /api-protect, /component-new, /page-new, /code-review, /docs-generate, /commit-best, /lint, /code-explain, /code-optimize, /code-cleanup, and more (16+ commands).
Agents (17): backend-architect, frontend-architect, backend-reviewer, frontend-reviewer, e2e-runner, database-expert, requirements-analyst, tech-stack-researcher, security-engineer, performance-engineer, architecture-strategist, system-architect, pattern-recognition-specialist, refactoring-expert, learning-guide, technical-writer, deep-research-agent—all aligned to the compounding cycle.
Skills (19): project-manager, feature-planning, agent-selection, code-review, api-design-patterns, api-testing, e2e-playwright, accessibility-checklist, backend-architect, frontend-architect, backend-reviewer, frontend-reviewer, security-audit, performance-profiling, refactoring-checklist, requirements-discovery, docs-structure, postgresql, nosql-databases.
Rules (6): core-standards, compounding-dev-cycle, typescript, react, api-routes, e2e-tests.
Best for
Next.js, TypeScript, React, Supabase, and full-stack projects. Works with any stack.
Requirements
Cursor, Claude Code, GitHub Copilot, or Codex (at least one). For Cursor, apply the workflow via VS Code first; then you get the full slash-command experience in Cursor.
Customization
Edit workflow content per assistant:
The Cursor plugin manifest is in workflow/cursor-plugin/ (paths inside point to .cursor/ when copied). Apply copies workflow/cursor/ and workflow/cursor-plugin/ to your workspace .cursor/ and .cursor-plugin/.
- Claude Code:
workflow/claude/ — agents, rules, skills, hooks, CLAUDE.md. Apply copies to .claude/ and CLAUDE.md.
- GitHub Copilot:
workflow/copilot/ — copilot-instructions.md, instructions/, agents/, workflows/, ISSUE_TEMPLATE/, AGENTS.md, user-instructions.md. Apply copies to .github/ and root AGENTS.md.
- Codex:
workflow/codex/ — skills/ (synced to ~/.codex/skills/), AGENTS.md (workspace), user/AGENTS.md (user-level). Apply syncs skills and writes both AGENTS.md files.
After changing any tree, rebuild and repackage the extension so the next Apply uses the updated content.
More info
- Developer guide: Command Palette → Plan-Code-Review: Open Developer Guide
- Publishing: Command Palette → Plan-Code-Review: Open Publishing Guide
License & author
Licensed under MIT. See LICENSE.
