oozoofrog/codex-skills/local-ai-studio

Local AI Studio

Complete the user's requested Local AI Studio or on-device model task. For LocalAIHub, install a missing prerequisite using setup guidance, then continue the same task. For local Whisper speech recognition, transcription or timestamp estimates, use the separate Whisper workflow; it does not require LocalAIHub installation. Do not start a large local installation for a generic media request. User-selected tools, models, output locations, and formats take precedence; do not silently substitute this CLI, a cloud service, or another model.

Discover before running

• Resolve the selected backend first: LocalAIHub uses ai below; Whisper uses whisper-cli and its own discovery and model checks. Do not invent a Whisper flag for ai transcribe or claim native app integration from a separate CLI run.
• Resolve the user's executable first, otherwise inspect command -v ai. Confirm it is Local AI Studio with ai --help before using it; an unrelated executable named ai is not sufficient. Read current top-level help for every new task; do not assume subcommand --help is supported.
• Find the selected storage root from AIHUB_ROOT, the app's configuration, and ai status; then inspect ai models. On the reference Mac, /Volumes/eyedisk/AI/bin/ai and /Volumes/eyedisk/AI/LocalAIHub/README.md are discovery candidates only, not defaults for other machines. If a configured volume is missing, preserve that choice rather than creating a replacement root.
• Prefer the installed CLI, matching LocalAIHub README and current code over the older parent AI README, especially for music and translation. Read command/output details for the selected modality. If LocalAIHub or the needed group is missing, follow setup guidance. If the fetched version lacks the requested command, report the version mismatch instead of claiming support.
• READY checks required local files, not successful inference, output quality, or even every runtime dependency. Check the relevant group's readiness without running an extra generation as a probe.

Execute the requested task

1. Resolve input files and the requested deliverable. Treat user-provided inputs as authorized for the requested operation unless there is a concrete reason to doubt it. Voice cloning needs an authorized speaker sample and its exact transcript. Keep inputs local. Check applicable model terms: the current Qwen Image 2.1 public license allows non-commercial research or evaluation; for commercial use, check whether the user has a separate license before running it.
2. Select the matching command from current help. Translation here is English → Korean, not arbitrary language pairs. ai music --prompt ... creates one song without a server; bare ai music opens a persistent Gradio UI, so use it only when a UI is requested.
3. Resolve an absolute output destination before launch. Preserve user paths/formats, check writable space and existing files, and avoid overwriting sources or prior outputs without permission. If unspecified, use the selected root's output folders and unique names. --output is a directory for TTS, a stem for transcription, and a file path for the other generators/translation. Unsupported formats need an agreed conversion, not a renamed extension.
4. On a 24 GB Mac, run only one large generator at a time, including other CLI/app/UI sessions. Check for active jobs; wait or coordinate rather than killing unrelated work. Keep a requested size/duration; where unspecified, start with a modest deliverable. Video --memory-mode parallel is an internal mode, not permission for concurrent model jobs.
5. Pass prompts, lyrics and paths as safely quoted arguments or an argument array, never shell-evaluated user text. Capture the command/settings, exit status and actual output locations; wait for completion. If a run fails, inspect the error and any partial output before a bounded retry. On OOM, missing runtime, or unresponsive execution, stop and report; do not loop, download/reinstall, or launch duplicates automatically.

Verify and deliver

• A printed Output: line is an intended destination emitted before inference, not proof of success. Require a completed run plus newly written, nonempty files tied to that run. For TTS, enumerate actual files in the chosen directory/prefix; do not hand back only a directory or assume a filename.
• Check actual file format/decodability, not just extension. Inspect image dimensions and preview; inspect audio/video duration, streams and playable samples; read UTF-8 translation/transcription and validate the requested subtitle/JSON structure. Compare content to the request where tools allow. Use verification examples as needed.
• Report readiness, process completion, file/format checks, and visual/listening/semantic review separately. If playback or viewing is unavailable, say so; never infer quality from READY or exit code alone.
• For transcription, distinguish recognized text, estimated alignment, human review and applied timing corrections. A token timestamp or agreement between same-model crops does not verify a sung onset/end. Preserve existing subtitles until the requested correction has adequate evidence.
• Deliver the verified files with usable paths/links (or real attachments when supported), key settings and any limitations. A path is not an attachment. Keep model weights, environments, generated media, private logs and credentials out of this skills repository.