Crun Agent Skills

Crun Media Agent Skills

Use this skill as the entry point for Crun media work. Resolve all commands from this directory:

• runtime/crun_cli.py handles model discovery, credit checks, task creation, status, polling, local-file upload, and generated-media downloads. Upload with crun_cli.py upload <local-file>.
• catalog/models.json is the local routing catalog.
• skills/ contains the core pipeline skills, and skills/scenarios/ the scenario skills, listed below; read the relevant child SKILL.md before performing that part of the workflow.

Select the workflow

Core pipeline skills — the shared infrastructure every request flows through (route → estimate → create/monitor):

Scenario skills (skills/scenarios/) — matched by user intent; each one composes the core pipeline internally:

For a broad end-to-end request — the common case where the user describes the media they want but does not know Crun model names or payloads — follow the "Orchestrate safely" steps below, reading each child skill as that step needs it. Treat crun-task-runner as the single authority for task creation, authorization gates, timeout recovery, failure handling, and result delivery. Read it before creating any task.

Pass JSON portably

Prefer a UTF-8 JSON file on Windows and whenever shell quoting is uncertain:

python <skill-root>/runtime/crun_cli.py task estimate --model <model> --input-file <input.json>
python <skill-root>/runtime/crun_cli.py models route --intent-file <intent.json>

Use - instead of a path to read that JSON object from stdin. Use --input-json or --intent-json only when the current shell can preserve the JSON string exactly. Each JSON source must contain one object.

Upload a new local input image, video, or audio file before constructing model input:

python <skill-root>/runtime/crun_cli.py upload <local-file>

Use the returned file_url. Never send Base64, data URIs, or local paths as Crun media inputs. For derivative work, reuse the resource URL returned by the earlier Crun task instead of downloading and uploading it again.

Orchestrate safely

1. Identify the output modality and operation. Collect only indispensable inputs, and require source media for edits, references, lip-sync, face-swap, watermark-removal, voice-cloning, and similar transformations.
2. Upload only new local source media (crun_cli.py upload); reuse an existing Crun resource URL directly.
3. Route only when the user did not specify an exact model (crun-model-router), then inspect the selected model's live schema and construct only supported input fields.
4. Estimate the exact model input before every new task (crun-account-credits). For a routed model, report the model, estimated credits, and relevant settings, then obtain explicit confirmation. Skip only this extra confirmation when the user explicitly named the model — never treat the original generation request as the confirmation.
5. Create once with task create, capture the returned task_id, and continue only with task wait or task status using that ID.
6. Follow crun-task-runner for all failures, timeouts, retries, and result delivery.

Use balanced quality and speed when the user has no preference. Ask about native audio only when it changes the viable models, and do not ask the user to choose a provider unless the alternatives have materially different tradeoffs.

Always estimate affordability before creating any task, including tasks for a user-specified model. Require affordable: true from task estimate before task create. The reason is that CreateTask charges credits and is never retried, so an unaffordable or malformed submission wastes a real charge that cannot be undone.

Keep task run and media run only for compatibility or deliberate one-shot use. They create a task directly, so estimate affordability yourself beforehand and confirm a routed model first — they do not gate on either. Do not use them as the default agent workflow.

Handle sensitive media

Before transforming a real person's face or voice, or removing a watermark, require confirmation that the user owns or is authorized to transform the source. Preserve any disclosure or labeling request. Refuse impersonation, fraud, non-consensual sexual content, and unauthorized removal of an ownership mark.

Keep API keys out of task payloads, output, and committed files. Let the runtime resolve CRUN_API_KEY in this order: the ~/.crun/.env file first, then the CRUN_API_KEY environment variable. When no key is configured, the runtime returns configuration_options — an ordered list with a permanent setup command for each method split by platform (macos_linux, windows_cmd, windows_powershell). Recommend the first option: the CLI's own python "<absolute path to runtime>/crun_cli.py" config set-api-key <your_api_key> command (already fully resolved in the payload), which validates the key and persists it into ~/.crun/.env so it works across sessions on every platform.