video-prompt-preflight-qa

name	video-prompt-preflight-qa
description	NexTide 视频生成前 Prompt QA。用于在消耗积分前检查短视频 prompt、分镜、爆款复刻请求或 AI 视频生成请求，识别弱开头、产品过早露出、viewer question 不清、漂移风险、缺少负向约束、UGC 真实感不足和输出格式错误。适合“生成前帮我检查”“这个提示词能不能跑”“别浪费积分先 QA 一下”。
allowed-tools	Read, Write

Video Prompt Preflight QA

Use this skill immediately before NexTide video/image generation or batch generation.

It is for:

• checking prompt quality before credits are spent
• catching weak openings and early product reveal
• identifying drift risk from vague prompts
• producing concise fix notes for rerun
• deciding whether a generation task can run now

It is not for:

• original hook creation → use short-video-hook-designer
• reference decoding → use reference-opening-decoder
• first-frame design → use visual-hook-optimizer

Core Rule

Judge prompt quality by controllability, not impressive prose.

Checklist

Check:

• opening strength
• first 1-3 seconds mechanism legibility
• viewer question clarity
• visible evidence / proof path
• promise-delivery match
• product reveal timing
• ad-detection risk
• negative constraints
• realism / UGC control
• identity/product consistency constraints
• output format correctness
• cost/risk if batch or high-cost generation

Required Output

Verdict:
Major risks:
- ...
Missing fields:
- ...
Likely drift:
- ...
Fix now:
- ...
Can run now:

Verdict values:

PASS
PASS_WITH_FIXES
RISKY_DO_NOT_RUN
UNDER_SPECIFIED

Blame Rule

Point weak prompts to the right upstream stage:

• bad route → short-video-hook-designer
• benchmark unclear → reference-opening-decoder
• frame one weak → visual-hook-optimizer
• generation payload missing required fields → target NexTide generation skill

NexTide Generation Guard

Before high/variable-cost runs, remind the user to use bounded inputs and cost confirmation when available:

--max-credits
--yes only after review

Never silently approve large batch generation with vague prompts.

References

• references/NOTICE.md

NexTide Capability Contract

视频提示词生成前 QA

• Capability: prompt.preflight.qa
• Version: 0.2.0
• Category: system
• Status: available
• Execution: local_agent
• Async: false
• Cost level: free
• Required auth: nexTideApiKey
• Required env: none
• Required plan: none
• Rate limit: none
• Estimated credits: 0
• Estimated duration: 60 seconds
• Tags: prompt-qa, preflight, video, cost-guard, local-agent

Description:

在消耗积分前检查短视频 prompt、分镜或 AI 视频生成请求，识别弱开头、产品过早露出、漂移风险和缺少约束。

Examples:

• 视频生成前 QA

  {
    "prompt": "生成一个 15 秒产品种草视频，开头展示便携榨汁杯，然后女生说它很好用。",
    "targetCapability": "viral.midform.video.generate"
  }
  

Input fields:

• prompt (string, required)：准备执行的视频/图片生成 prompt 或 storyboard。
• targetCapability (string)：计划调用的 NexTide capability，例如 viral.midform.video.generate。

Output fields:

• verdict (string)：PASS / PASS_WITH_FIXES / RISKY_DO_NOT_RUN / UNDER_SPECIFIED。
• fixes (array)：必须修复的问题。

CLI:

nextide capability run prompt.preflight.qa \
  --input .nextide/input/prompt.preflight.qa.json \
  --output .nextide/output/prompt.preflight.qa-result.json \
  --mode wait

If the result contains artifacts, export them:

RUN_ID=$(node -e "const r=require('./.nextide/output/prompt.preflight.qa-result.json'); console.log(r.run && r.run.runId)")
nextide run artifacts "$RUN_ID" \
  --output-dir .nextide/output/$RUN_ID \
  --download \
  --gallery \
  --datatable

Then read summary.json first, followed by manifest.json.

General Rules

• Use NexTide capability IDs, not internal n8n webhook URLs.
• Do not expose API secrets or internal webhook URLs in prompts or output.
• If status is not available, fail fast and explain what is missing.
• For async tasks, prefer --wait when the user wants a finished result in the same turn.
• After a finished run, use nextide run artifacts <run-id> --output-dir .nextide/output/<run-id> --download --gallery --datatable and read summary.json then manifest.json.
• For long-running runs, prefer nextide run follow <run-id> --output-dir .nextide/output/<run-id> --timeout 1800 --interval 5.
• Prefer returning summary.recommendedResponse.message, preview.html, datatable.json, and local artifact paths over pasting huge raw JSON.
• If the CLI output includes explanation, convert it into a clear user-facing failure message with next actions.