> ```{eval-rst} .. meta:: :description: Business-objective prompts that turn a plain-language manufacturing quality goal into a running Intel Edge AI application. ``` # Manufacturing Application Builder
The Manufacturing AI Suite **Manufacturing Application Builder** is a collection of reusable, business-objective prompts. Each prompt turns a plain-language goal into a running Intel® Edge AI application. An example goal is "I want to be alerted before a pump fails." Each prompt states only a business outcome. It does not name a framework, model, precision, or device. The prompt hands off to the `manufacturing-ai-app-builder` orchestrator skill. The orchestrator asks a few business questions and selects the right delegate skill. It then proposes a plan. After you confirm the plan, it builds the deliverable on Intel hardware. --- ## How prompt-driven application development works A prompt is intentionally minimal. It describes *what you want to achieve* and delegates every technical decision to the orchestrator skill. 1. **State a business objective.** Select the prompt that matches your outcome, for example "predict a pump failure" or "catch defective welds". Alternatively, describe your goal in your own words. 2. **The orchestrator takes over.** The [`manufacturing-ai-app-builder`](https://github.com/open-edge-platform/manufacturing-ai-suite/tree/release-2026.3.0/.github/skills/manufacturing-ai-app-builder) skill runs a short business Q&A. It asks for: - the outcome you want - what feeds the decision: a camera, a sensor, or both - your inputs, such as an RTSP camera, an OPC-UA server, an MQTT broker, or sample data - where the application runs and on which Intel hardware - how you want to be notified, for example a dashboard, MQTT alerts, or OpenMES issues - where to build the application on your host If your decision combines camera and sensor signals, the orchestrator also asks which deployment you want. The options are `multimodal` (fused alert only), `vllm` (adds written explanations from a vision-language model), or `agentic` (adds agent-coordinated analysis). The last two options need extra models and an Intel GPU, so the orchestrator asks you to opt in. 3. **Skill discovery.** From your answers, the orchestrator selects the delegate skill: | What feeds the decision | Delegate skill | | --- | --- | | A sensor stream only | `manufacturing-app-recipe` in `ts` mode | | A camera and sensors, combined into one verdict | `manufacturing-app-recipe` in `multimodal`, `vllm`, or `agentic` mode | | A camera only | `metro-ai-app-recipe`, which the orchestrator installs on demand | 4. **Plan and confirm.** The orchestrator presents the deliverable, the skills that build it, the technology it inferred, and the requirements. It then waits for your approval. Nothing is created before you confirm. 5. **Build by delegation.** After you approve, the orchestrator builds the deliverable by delegating to the chosen skills, and then verifies the result against each delegate's completion criteria. --- ## Before you begin The orchestrator is a skill that runs in an AI coding agent. You need: - An AI agent tool that supports skills, such as GitHub Copilot in Visual Studio Code. - Node.js 22.20 or later and internet access to install skills with `npx`. After you approve the plan, the orchestrator installs any missing delegate skill, including `manufacturing-app-recipe`, with `npx skills`. You don't need to clone this repository. - Docker with Docker Compose v2 on the host where the application runs. The orchestrator lists any further requirements, such as free ports and Intel GPU access, in the plan. --- ## Writing your own prompt You don't need to learn any special format. Describe the outcome you want in plain language and hand off to the `manufacturing-ai-app-builder` skill. Copy the following template into your AI agent tool, replace the objective with your own, and run it: ```text I want toLoading prompts…