A short‑form video script pipeline that runs on a schedule and needs no LLM API key

说明与用途

一个按计划运行且不需要LLM API密钥的短视频脚本生成管道，提供结构化JSON输出、平衡的类别预算，并过滤薄弱RSS源。

问题

Automated faceless YouTube/TikTok script generation pipelines often produce unstructured, inaccurate, or wasteful outputs due to generic prompts, uneven category budgets, thin RSS sources, and lack of structured timing.

结果

The pipeline produces reliable, segment‑structured short‑form video scripts with category‑specific hooks, balanced category budgets, and filtered content, reducing hallucinations and token waste.

工作流

• 1. Define mode (rss or direct) and input parameters (categories, maxItems, duration).
• 2. For rss mode, fetch items from each category feed, limiting to maxItems per category and interleaving results.
• 3. For each item, enforce a content‑length floor (~500 chars summary or 200+ body chars); skip items that do not meet the threshold.
• 4. Construct a prompt that includes the category, topic, and data, and request a structured JSON output with segments, speech, visual, start/end, and duration fields.
• 5. Send the prompt to an LLM without requiring an API key (e.g., local model or open‑source).
• 6. Validate the JSON structure; if parsing fails, retry or flag the item.
• 7. Store the generated script with metadata (category, topic, duration, wordCount).
• 8. Use the planned start/end times for downstream audio generation, captioning, or stock footage matching.
• 9. Monitor actual audio timing (actualStart/actualEnd) to debug timing mismatches.
• 10. Iterate and refine prompts per niche to improve hook quality.
• 11. Schedule the actor to run at desired intervals.
• 12. Export or publish the scripts to the video production workflow.

证据

The model returns prose with no segment structure roughly one time in six, and your downstream parser explodes.

来源

Thin sources produce hallucinated scripts… require a summary of at least ~500 characters, or 200+ characters of extracted body text, and skip the item otherwise.

来源