Skip to content

Mini App Build.md

Latest commit

 

History

History
77 lines (49 loc) · 3.1 KB

Mini App Build.md

File metadata and controls

77 lines (49 loc) · 3.1 KB
Raw

Mini App Build

Generate, evaluate, and self-evolve a complete mini-application from a single text prompt. The AWorld Agent orchestrates the full flow: automatic creation → automatic evaluation → automatic optimization. No manual intervention is required once you submit your prompt.

What You Get

  • A runnable mini-application (e.g. web app) generated from your description
  • Automatic evaluation and self-evolution until your quality criteria are met
  • One command from prompt to product; the entire loop runs autonomously

Example: English Word Learning App

This recipe reproduces the English word learning app flow shown in the main AWorld CLI examples.

User Prompt

Use this prompt (or adapt the goal and score threshold to your needs):

help me create an English word learning app, with a UI quality score over 0.9

How Quality Is Measured: Evaluator & Official Skill

The Evaluator agent scores the generated app using objective criteria defined by a Skill. For this recipe we use the official AWorld Skill:

  • Skill: app_evaluator (shipped with the repo)

If you have already installed AWorld and AWorld-CLI, the Evaluator loads this official skill by default. You do not need to install or configure the skill separately for this flow.

What Happens (Fully Automated)

  1. Create — The Developer agent turns your prompt into a runnable app (e.g. HTML/JS).
  2. Evaluate — The Evaluator runs the app_evaluator skill (e.g. UI quality score).
  3. Optimize — If the score is below your target (e.g. 0.9), AWorld instructs the Developer to fix issues; the loop repeats until the criteria are met.

All steps run automatically; you only need to start the CLI and submit the prompt.

How to Run

  1. Install and configure AWorld CLI

  2. Launch the CLI
    In your working directory, run:

aworld-cli

  1. Enter the app-creation flow
    Choose or enter the flow that creates an app from a prompt.

  2. Submit the prompt
    Paste or type:

help me create an English word learning app, with a UI quality score over 0.9
  1. Let it run
    The agent will generate the app, evaluate it with the official Evaluator skill, and iterate until the UI quality score meets your target. When done, run or deploy the output as needed.

For a reusable workspace setup, keep the same .env or interactive CLI configuration across runs.

Inspecting The Evaluator Report

When you want a stable machine-readable artifact for CI or post-processing, use the standalone evaluator command against the generated app or artifact:

aworld-cli evaluator --target ./artifact --output ./report.json
aworld-cli evaluator --print-report-schema
aworld-cli evaluator --validate-report ./report.json

The emitted report includes report_format, normalized metrics, structured gate, and automation fields. That makes it suitable for quality gates, regression checks, or downstream dashboards without parsing freeform evaluator text.