A development workflow for Claude Code that takes an idea to committed code: it interviews you, writes a reviewed plan, builds it task by task and remembers what it learned.
Coding agents tend to jump straight into code and fill the gaps with guesses. viber writes no code until you approve a plan that a reviewer has already checked against your codebase.
Questions come in short rounds, each decision with options and a recommendation. Add --prove and every recommendation is checked against your code before you see it.
The plan is reviewed before you approve it, each task before its commit, and the whole build before it ends.
A failed task retries up to five times, one model tier up each time. Where a build would stop to ask, an arbiter agent picks a way forward and records why.
Each task gets the cheapest model that can do it, between a floor and a ceiling you set, so your usage limits last longer.
Conventions go to .claude/rules/, area knowledge to that area's CLAUDE.md. You explain less with every run.
The plan and its progress are committed, so an interrupted build resumes in a new session or on another machine.
Build and test commands and conventions come from your own CLAUDE.md and .claude/rules/, whatever language you work in.
Beyond the core loop, viber talks a topic through with you against trusted sources, triage's issues,
opens GitHub issues and pull requests, mocks up UI changes, holds every interface it builds to
professional design standards, writes Playwright tests, audits large codebase's, checks whether a
side-project idea is worth building, writes and tunes your own Claude Code skills and agents for
the model they run on, and hands a session over to the next one.
The viber guide lists every command.
You describe what you want with /viber:intent. viber asks in short rounds, every decision with
concrete options and a recommendation, and says so when an answer is weak. It saves your confirmed
answers as a summary file for the planner, which turns them into a plan, and a reviewer checks that plan against your code before you see it. Approve it and
walk away: viber builds the plan task by task, commits each task and finishes on your own test
suite. Then it writes down what it learned, so the next run starts from there.
| Step | You run | You get |
|---|---|---|
| Understand | /viber:intent, or /viber:fixer for a bug |
A confirmed summary of the change |
| Plan | nothing, it follows the interview | A plan reviewed against your code, waiting for your yes |
| Build | nothing, it starts on your yes | One commit per task, a final review, a green test run |
| Remember | nothing, it closes the build | Updated CLAUDE.md files and .claude/rules/ |
claude plugin marketplace add https://github.com/p2p2sp/viber --scope user
claude plugin install viber@p2p2 --scope user
Then, in your project:
/viber:setuponce: it writes the config, the ignore rules and the permissions./viber:intentto start your first change.
/viber:help opens the full usage guide in your browser. --scope user makes the plugin available
in all your projects; drop it to install for the current repository only.
- viber runs in Claude Code only.
- A plan written outside the planner is only reviewed, through
planning.outside-plan-review. - Issue and pull request steps work with GitHub, through the
ghCLI. - It changes fast, and a release can change how it behaves. Run
/viber:setupagain after an upgrade.
| Plugin | Needs |
|---|---|
| viber | the gh CLI for its GitHub steps; optional: node for /viber:setup's settings merge, Node.js 22.6 or newer for pro-designer's contrast check, Node.js 18 or newer and web access for /viber:idea-validator, Playwright for /viber:e2e |
These plugins are opinionated and their maintainers set the direction, so pull requests are not accepted. Bug reports and ideas are welcome: open an issue. More in CONTRIBUTING.md.
Built and maintained by P2P2.