tfcbot/agent-video-team · Archived

director-qa

>- Run Vidjutsu QA gates on completed clone videos: anatomy frames, structured Watch criticism, transcript verification, and explicitly enabled VidLang rules.

First seen Apr 12, 2026

Installation

$ npx skills add tfcbot/agent-video-team --skill director-qa

Stronger alternatives

This repository is archived — consider an actively maintained alternative.

Similar popular skills

Related neighbors and high-traction skills in the same topics — useful to compare before installing.

Also in this package

Other skills from tfcbot/agent-video-team · top by installs.

npx skills add tfcbot/agent-video-team

Browse all from tfcbot/agent-video-team

More details

Agent compatibility

Declared targets from SKILL.md / docs. Unmarked agents are not listed — the skill may still install via the CLI.

Claude Code Not declared
Cursor Not declared
Codex Not declared
GitHub Copilot Not declared
Windsurf Not declared
Gemini CLI Not declared
Cline Not declared
OpenCode Not declared

Repository health

License LICENSE
Default branch main
Open issues 0
Status Archived

Package contents

Files included with this skill beyond the listing page.

  • skill md SKILL.md 2,934 B
  • docs SUMMARY.md 174 B

History

  1. First seen on skills.sh
  2. First recorded snapshot · 2 installs

SKILL.md

Director — QA

Run each gate in order. If a gate fails, revise the motion prompt and submit a new Vidjutsu clone only after the prior task is terminal. Allow at most five total generation attempts.

Gate 1 — Anatomy

Extract three frames from the completed clone:

vidjutsu extract --mediaUrl "$CLIP_URL" --frames auto

For each returned frame URL, request an explicit response schema:

vidjutsu watch --mediaUrl "$FRAME_URL" --prompt 'Return only JSON: {"score":number,"issues":[string]}. Trace every visible limb to the torso and check joint position, body-part count, and spatial continuity. Score 8+ only when anatomy is plausible.'

Read response.score. Reject if any frame scores below 8. Extract is limited to 100/day; Watch is limited to 50/day.

Gate 2 — Visual critic

vidjutsu watch --mediaUrl "$CLIP_URL" --prompt 'Return only JSON: {"score":number,"issues":[{"severity":"minor"|"major"|"critical","category":string,"description":string}]}. Check identity continuity, anatomy, artifacts, motion, framing, and audio sync.'

Read the result from the API's response envelope. Reject for a score below 8 or any major/critical issue. Adapt the prompt to the content style.

Gate 3 — Speech

Skip when there is no spoken dialogue.

vidjutsu transcribe --mediaUrl "$CLIP_URL"

Compare the returned transcript with intended dialogue. Reject missing, incorrect, or repeated phrases. Number-word normalization is non-blocking. Vidjutsu transcribe is limited to 30/day. This gate verifies the original Kling soundtrack; Vidjutsu does not replace voices.

Gate 4 — VidLang rules

Every POST /v1/check request requires both spec and rules, and at least one rule must be enabled. Rules are off by default; an omitted, empty, or all-disabled rules object is invalid.

vidjutsu check run --spec ./spec.json --rules '{"VL013":true,"VL011":{"severity":"warning"}}'

Use true to enable a rule, false or { "enabled": false } to disable it, and { "severity": "warning" } to enable with an override.

Custom plain-text rules are stored separately:

vidjutsu check rules list
vidjutsu check rules set 'Dialogue contains the AI reveal,Requested CTA is present'

GET and PUT /v1/check/rules are unmetered. Evaluate those strings in the agent layer; Vidjutsu stores but does not execute them. Built-in checks are limited to 100/day.

Finish

For each passing clip, report its URL, Watch scores, enabled VidLang rules, and attempt count. To chain another scene, extract the final frame with:

vidjutsu extract --mediaUrl "$CLIP_URL" --frames last

Do not continue to post-production if any required gate fails.