VibeWise: you design it, Claude writes it

Ask for a feature with VibeWise on and the first thing back is a question: how would you build it? You make the design calls, the agent explains concepts and pushes on weak spots, and code gets written only once you approve a named step. Noah Kim publishes the plugin under the MIT license for Claude Code and Codex. Anthropic doesn't publish it, though the README installs it from Claude Code's built-in plugin directory. For how plugins and skills relate, see Claude Code skills.

The three checkpoints

• Build. You reason through how to approach the problem. One focused question can invite a whole approach.
• Design. The agent sums up the design and its trade-offs. "Confirm and continue" records it; no code is written yet.
• Implementation. The agent names the exact changes it's ready to make, and "Implement this step" authorizes only that scope.

They aren't three fixed stops: an Implementation checkpoint can confirm the design too. Every confirmation also offers Discuss. After coding, the agent reports what changed, which tests it added and which checks ran.

The rules in behavior.md stop the agent steering you towards its own plan. It offers options only when you ask or get stuck, and a slow or short answer isn't treated as being stuck. It still flags concrete errors and risks.

Levels and saved notes

Setup asks one question at a time: new project or existing repo, your programming and stack experience, and whether to keep defaults. Beginners get more explanation and diagrams; advanced users get harder questions about failure modes. Checkpoint frequency (Light, Normal or Frequent) is a separate setting. On an existing repo the agent inspects the code and sketches a small system map before anything else.

Your profile, progress and that map live in .vibe-wise/ at the repository root. A session-start hook reloads them after a restart or a compaction, so learning mode survives both.

How to install it

You need a current Claude Code or Codex and Python 3, with no extra packages. In Claude Code:

/plugin install vibe-wise@anthropic-plugin-directory

Restart Claude Code in your project and run /vibe-wise:learn. If the directory isn't in your version, add the GitHub marketplace instead with /plugin marketplace add nykooi1/vibe-wise, then /plugin install vibe-wise@vibe-wise. For Codex it's codex plugin marketplace add nykooi1/vibe-wise and codex plugin add vibe-wise@vibe-wise, then $vibe-wise:learn.

Example request

/vibe-wise:learn
I want to build [what, e.g. a notes app where one note can sit in several folders] with [stack]. I'm [a beginner | intermediate | advanced]. Before you write any code, ask me how I'd handle [the hard part, e.g. deleting a folder that shares notes with another].

Mid-session you can say "use fewer checkpoints", "just implement this one" or "pause learning".

Limits

• Slower by design. The behaviour rules put learning and your control ahead of build speed, so a feature takes more turns than plain Claude Code.
• Codex needs a nudge. Run $vibe-wise:learn at the start of each Codex session and whenever it seems to lose track.
• Notes enter the context. There's no account, backend or telemetry, but saved notes are sent to the model like any other context, under your normal Claude Code or Codex data settings.
• Git is up to you. The plugin won't add .vibe-wise/ to .gitignore silently; add it yourself to keep notes out of commits.
• Reset backs up first. /vibe-wise:reset asks Cancel or Reset learning, then saves the old notes under backups/ and leaves your source code alone.

When to use it

It fits a student, a career changer or a developer in an unfamiliar stack who wants to finish a project and still be able to explain it. To learn an existing codebase by re-deriving its decisions, try Rederive. To have a loose plan questioned before any build, use Grill Me.