For "vibe coders" —developers who flow and build using prompts and AI assistance in tools like Cursor, Windsurf, or Cline— YapuCli is the perfect exoskeleton. Its main goal is to lift the mental burden of organizing context, allowing you to focus purely on creativity and business logic (the vibe).
Below is the daily development cycle using the full suite of Yapu commands.
Tip
Each workflow phase is designed to be executed by your AI. Your role is to supervise, validate, and adjust — not write code manually.
When starting a new project or feature, instead of wasting an hour writing excessively long prompts so the AI understands your architecture rules, you simply open the terminal in your empty directory and run:
npx yapu-cli init
Yapu founds the "colony" by creating the directory structure in .planning/. It automatically clones the commandments learned from previous projects (Hive Mind) and leaves the ROADMAP.md and STATE.md files ready in the root.
You don't have to explain anything to your AI. You simply tell your agent: "Read the PROJECT.md, analyze my initial idea, and structure Phase 1 in the ROADMAP.md". The AI will know exactly where and how to write.
As your agent starts writing code, reading files, and proposing changes, you can easily lose track of exactly what task the AI is working on or what percentage of the phase is completed.
Open a secondary terminal (ideally on an attached monitor) and run:
yapu dash
Yapu draws an interactive dashboard (TUI) without using any external dependencies. This panel is divided into logical areas:
- Top: Progress bar based on the checkboxes in
ROADMAP.md. - Left: Pending and completed tasks extracted from
STATE.md. - Right: Live activity logs intercepted from the AI's brain.
You feel the progress flowing visually as if you were in the matrix. It allows you to supervise your agent at a quick glance without having to manually open markdown files to see what was crossed out.
You've pushed your code to the main branch and stepped away for a coffee. However, your CI/CD pipeline (GitHub Actions) fails because an environment variable was missing or an asynchronous test wasn't handled properly.
The pipeline has the yapu-heal.yml flow pre-installed. Upon detecting the failure in npm test, it dumps the error and automatically runs yapu rescue error.log. This immediately creates a rescue session in .planning/debug/RESCUE_YYYY-MM-DD.md with the fully formatted log and instructions to fix it.
You return with your coffee and receive the failure notification. Instead of manually debugging, you tell your AI: "Read the rescue task generated by Yapu and fix it". The AI reads the error, injects the patch, and you push again without losing your creative flow.
Your new feature passes all unit tests. Everything seems perfect, but you wonder: Will it survive if the database gets slow or if the JWT expires mid-session?
Instead of writing complex integration or infrastructure tests, you summon the chaos.
You command your AI agent: "Activate chaos mode. Execute the directives in .agents/skills/yapu-chaos.md".
The AI becomes a Chaos Monkey. It will start injecting artificial latencies (e.g., a 5s setTimeout in your routes), forcing unhandled exceptions, or deliberately dropping promises in your local environment. If your Node.js server explodes from an Unhandled Rejection, the AI will document the crash and automatically write the necessary try/catch blocks or graceful degradations so the system becomes resilient.
Weeks of intense development have passed. You've closed multiple phases. You notice that Cursor, Windsurf, or Cline is getting slow, expensive in token consumption, or starting to "hallucinate" because there are dozens of SUMMARY.md files and old tasks overwhelming its context window.
It's time to clean up the project's memory:
yapu gc
YapuCli scans .planning/phases/. It takes all old phases, concatenates their raw text into a temporary file, and archives (moves) the original files to a hidden .planning/archive/YYYY-MM-DD/ folder, removing them from the main tree so the AI stops indexing them. Finally, it generates a specialized compression task (LORE_MASTER.md).
You just have to tell your AI: "Execute .planning/tasks/LORE_MASTER.md". The AI will process weeks of history and condense it all into a highly dense document called LORE.md (retaining only architectural decisions and the why of things, removing noise like trivial tasks). You immediately regain your AI's speed, lower token costs, and keep a clean repository.