How to use the KrillinAI Pipeline agent skill

What the skill makes the agent do

Agent skills are instruction packages that an AI agent loads when a task matches their description. This one is about planning and checking KrillinAI jobs. The agent maps the outputs you ask for to KrillinAI stages, validates the plan with a dry-run, and then runs each stage command separately.

The skill does not run the whole pipeline in one go. The source states that non-dry-run pipeline execution returns unsupported_command, so the agent must run real work through the individual stage commands.

• Supported outputs: subtitle, tts (dubbing), horizontal-bilingual, horizontal-dubbed, vertical-bilingual, vertical-dubbed, and cover.
• The dry-run validates output names and returns a pipeline-stage JSON response. It does not execute stages or write a task manifest.

Install the skill

Option 1: download the ZIP from the skill page and unzip it into ~/.claude/skills/ for your personal skills, or into a project's .claude/skills/ folder to scope it to one project. The package contains a single file, SKILL.md.

Option 2: in the Claude apps, open Settings → Capabilities → Skills and upload the ZIP.

Set up the requirements

The skill only orchestrates an existing tool. You need the KrillinAI CLI installed before anything else. The skill finds it through the KRILLINAI_CLI environment variable, and uses KRILLINAI_CWD to point to a working directory.

Stage commands also need an absolute --workdir path. Provider configuration for TTS and other stages is required, but it is defined in the shared krillinai-cli contract, which is not included in this skill. The source points to skills/krillinai-cli/references/cli-contract.md for those details.

• KrillinAI CLI installed and reachable through KRILLINAI_CLI.
• KRILLINAI_CWD set to a working directory.
• An absolute --workdir path for each stage command.
• Provider configuration for TTS and other stages, per the shared CLI contract.

Trigger the skill with natural requests

You do not need special syntax. Describe the result you want and mention the video work. The skill is a good match when you want a plan checked, documented, or verified for a multi-stage KrillinAI output.

Useful requests name the outputs you need, for example a subtitled file plus a portrait bilingual video. Ask the agent to validate the plan first if you want a check before any processing starts.

• Plan a KrillinAI job that makes subtitles, dubbing and a vertical bilingual video, and dry-run it first.
• Check whether these outputs are valid for KrillinAI: subtitle, tts, horizontal-dubbed, cover.
• After the stages finish, confirm the requested video and cover files exist and match the plan.

Run a dry-run plan check

The dry-run is the safe first step. Replace the output list with the outputs you need. The command runs from your KRILLINAI_CWD and calls the CLI with the --dry-run flag.

A successful check returns terminal JSON with "ok": true. The dry-run only validates names; it does no real work.

Execute the plan stage by stage

After a valid dry-run, the agent maps each output to a stage command: subtitle, tts, render-horizontal, render-vertical, and cover. Each command runs independently, and all of them use the same absolute --workdir.

If something fails, use the recovery approach below rather than starting over. Read krillinai_manifest.json before rerunning, so you do not repeat expensive work that already succeeded.

• Map outputs: subtitle → subtitle, tts → tts, horizontal-* and vertical-* → render-horizontal or render-vertical, cover → cover.
• Keep the same absolute --workdir for every stage.

Recover from failures

Recovery is stage-specific. If subtitles succeed but a render fails, rerun only the render with the same --workdir. If TTS fails, keep the subtitle outputs, fix the provider configuration, and rerun tts.

Always check krillinai_manifest.json first to see which stages are already complete.

• Render failed after subtitles succeeded: rerun only the render stage with the same --workdir.
• TTS failed: keep subtitle outputs, fix provider config, rerun tts.
• Before any rerun: read krillinai_manifest.json.

Verify the results

A stage reporting success is not the end of the check. The source asks the agent to confirm manifest stages and the requested output files, not just the plan response.

Open the final media and cover files you asked for, not only the intermediate outputs.

• Dry-run: require terminal JSON with "ok": true.
• Real work: confirm every independently executed stage succeeded.
• Confirm manifest stages and requested output files exist.
• Inspect the final requested media and cover files.

Safety notes: scripts and network

This skill runs scripts and uses the network. The source does not describe exactly what data is sent to speech or translation providers, so review that before processing sensitive material. Keep sensitive content out of workflows until you have checked the provider setup.

The skill is an orchestration layer over the KrillinAI CLI. Only run it on a machine and project where you trust the CLI and its configuration.

• Scripts run on your machine through the CLI commands shown above.
• Network use is involved, including provider calls for TTS and other stages.
• The skill does not state what media leaves the machine.