caption-local/docs/AI-SETUP.md

Set up Caption Local with an AI assistant

All guides · Install · Troubleshooting

Using an AI assistant is optional. You can install the service yourself using the README. The service itself does not require an AI account or API key.

Copy-and-paste prompt

Give an assistant with terminal access this prompt:

Install Caption Local from https://github.com/steveseguin/caption-local on this computer. Read README.md, DEPLOYMENT.md, OPERATIONS.md and skills/deploy-caption-local/SKILL.md. Preserve any existing installation and model cache. Use the latest tagged release unless I request development code. Choose native Python or Docker based on what is already available. Default to CPU and the multilingual small model. Keep the service bound to localhost and caption.ninja sharing off. Install dependencies, download the model, start the service and verify readiness plus real transcription. Explain how to open the capture page, connect a microphone, stop the service, and later update it. Report what you actually tested and any unvalidated Windows/GPU capabilities. Do not expose a public port.

Add your preferences as needed:

• “Use Docker and keep it running after I close the terminal.”
• “Inference runs on a Linux server; my microphone is on another computer. Set up SSH forwarding.”
• “I need twelve English streams. Evaluate the CPU throughput preset and measure capacity on this machine.”
• “I need Spanish transcription and English translation. Keep small and test accuracy.”
• “I have an NVIDIA GPU. Verify the actual driver/runtime and inference before claiming acceleration works.”

Reusable deployment skill

The folder skills/deploy-caption-local contains a plain-Markdown deployment skill. Ask your assistant to read that file, or copy the whole folder into the user skill directory supported by your assistant. Instructions for skill installation depend on the assistant; reading the file directly works without a special installer.

The skill routes installation details to the maintained guides instead of keeping separate copies of commands. It covers native/Docker setup, offline models, SSH, CPU/GPU choices and honest validation of stream capacity.

Check the result

The assistant should report the selected version, model, actual device and startup command; show that /health is ready and a real audio request works; and explain shutdown and recovery. A successful import, a detected GPU, or a mocked test alone does not establish a working transcription deployment.