A project-governance Skill for Codex and Claude Code that keeps AI-assisted coding sessions aligned with the user's intent, active phase, and explicit non-goals.
AI coding agents are strong at implementation, but long sessions often drift:
- the original plan gets forgotten
- new features appear without approval
- vague ideas become oversized builds
- implementation continues after a phase should have stopped
- review happens only after the project has already expanded
I created this Skill after running multi-step projects with Claude Code and Codex where the agent would start with a reasonable plan, then later lose track of that plan or add features that were never approved. The result was not a lack of code, but too much ungoverned code.
The useful pattern was clear: the project needed a lightweight supervisor workflow. Before coding, the agent should clarify the idea. Before implementation, it should produce a phase plan. During implementation, it should stay inside the active phase. When scope changes, it should stop and create a new phase instead of expanding silently.
Vibe Coding Review packages that workflow into a reusable Skill.
- Turns vague project ideas into clearer build directions
- Requires project anchors before implementation
- Reviews Claude Code or Codex plans before coding
- Supervises active development phases
- Detects scope drift and unapproved feature expansion
- Structures targeted Codex review handoffs
- Supports project closeout before publishing, deployment, or presentation
- Initial Setup / Anchor Creation
- Brainstorming Ask
- Plan Review
- Phase Supervisor
- Diff Review
- Scope Change / Phase 2 Plan
- Project Closeout
Governed projects should keep these files:
docs/
PROJECT_BRIEF.md
phases/
ACTIVE_PHASE.md
PROJECT_BRIEF.md captures durable project intent, boundaries, and non-goals.
ACTIVE_PHASE.md captures the current phase contract: objective, allowed scope, review gates, out-of-scope items, and handoff format.
For larger projects, optional phase files can be added under docs/phases/.
Copy this folder to:
~/.codex/skills/vibe-coding-review/
Copy this folder to:
~/.claude/skills/vibe-coding-review/
Restart the app or open a new session if the Skill does not appear immediately.
Explicit invocation:
Use $vibe-coding-review to brainstorm this project idea before planning.
Plan review:
Use $vibe-coding-review.
Here is Claude Code's plan. Review whether it fits PROJECT_BRIEF.md and ACTIVE_PHASE.md before I approve implementation.
Scope change:
Use $vibe-coding-review.
The active phase only allows README edits, but I now want to add a CLI and GitHub publishing. Tell me whether this fits the current phase.
- Start with Brainstorming Ask.
- Create or review project anchors.
- Produce a file-level phase plan.
- Send the plan to Codex for review when required.
- Implement only the approved active phase.
- Run Diff Review.
- Close out the phase before expanding scope.
- Not a code-generation framework
- Not a universal architecture guide
- Not a replacement for tests
- Not an automation runner
- Not a GitHub publishing workflow
- Not permission to expand scope without a new phase
SKILL.md
agents/
openai.yaml
references/
codex-review-gates.md
handoff-examples.md
lifecycle-mode-details.md
project-anchor-templates.md
- Low freedom for lifecycle gates and output contracts
- High freedom for brainstorming questions
- Explicit non-goals before implementation
- New feature requests become new phase plans
- Codex review prompts should be targeted, not vague
This is an initial public Skill release. It is intentionally small and procedural: the goal is to prevent AI-assisted project drift, not to replace engineering judgment.