ra-video-title

Video Title Skill

Core Principles

Define the topic first, then create the title. A video title must serve the core proposition of this video, and must not merely apply generic hooks like "AI tool", "flopped", or "found the reason".

Default replacement for dbs-xhs-title handling video titles.dbs-xhs-title It can still be used for general Xiaohongshu image-text titles or pure formula browsing; finished videos, video rewrites, benchmark titles, and two-part Douyin titles all go through this Skill.

Position in the two-stage workflow

In the two-stage workflow of "rewriting (Claude) → to-be-produced queue → production (Codex)", this skill runs during therewriting stage, and the input is the just-rewritten final script (the moment when the topic is clearest):

• Only produce candidates: 8-12 two-part candidates + Top 3 recommendation reasons, written into the "Title Candidates" section of the handoff draft
• Do not finalize during the rewriting stage: the final title is manually selected by the user from the candidates at publishing time
• The production side is unrelated to the published title: do not create titles, do not apply titles, do not generate 视频标题.md

Workflow

1. Lock the video topic

   • Prioritize reading the final script provided by the user, the video production brief,narration_segments.json, project description, or finished-video directory context.
   • Output a one-sentence topic judgment, in the format:这条视频讲的是：...
   • Title candidates must be able to tie back to this topic sentence.

2. Break down benchmark titles

   • If the user provides benchmark titles, extract commonalities: structure, first-half hook, second-half supplement, emotion, professional terms, result promise.
   • Do not copy the benchmark's specific expressions, and do not write another account's persona into the title.
   • By default, focus on checking the "two-part" form: the first half raises a scene/conflict/problem, and the second half gives the result/explanation/suspense.

3. Generate candidates

   • Generate 8-12 candidates.
   • By default use the two-part form, with common connectors:，、？、！、：。
   • Each candidate must simultaneously satisfy:
     • Hits the video topic, not a generic AI title;
     • The first half has a hook;
     • The second half supplements the direction or suspense;
     • Colloquial, suitable for short-video platforms;
     • Does not fully reveal the video's answer;
     • For news-type topics (new paper/new product/new event): the first half directly gives the factual subject (who + released what), without using metaphor/personification as the primary hook; metaphor can appear at most in the second-half supplement.

4. Filter and recommend

   • Give Top 3, and explain each with one sentence.
   • If the user asks "you decide", "write it directly", or "choose for me", select 1 final title.
   • If the user explicitly says "I'll choose", only give candidates and wait for a number.

5. Write to file

   • In the two-stage rewriting workflow: the only write target is the "Title Candidates" section of the handoff draft (8-12 candidates + Top 3 recommendations), do not write frontmatter, do not generate 视频标题.md, and the final selection is made manually by the user.
   • In other scenarios, when the user requests writing, in the location of the final MP4's 成片/ Create or update the table of contents 视频标题.md。
   • Write only the final title on the first line of the file.
   • Only append when the user requests retention or the candidates are valuable for later use ## 候选标题 section.
   • Use apply_patch to edit 视频标题.md. Do not modify the MP4 or other final video files.

Title scoring

Run each candidate through the following 5 points in your head:

• Topic fit: the closer to the video topic, the better; deduct points for titles that could apply to any AI video.
• Two-part tension: the two parts should build on each other, not simply repeat.
• Reason to click: why would the user click? Is it fear of falling behind, fear of pitfalls, wanting to save effort, wanting to get stronger, or wanting to see controversy?
• Information gap: the title should leave a gap; do not state the conclusion fully.
• Conversational feel: like a Douyin title, not an academic title or course title.

Prohibited items

• Do not come up with a title directly without reading the topic.
• Do not use titles that start with metaphor/personification for news-type videos (e.g. "Tell AI not to think about the Golden Gate Bridge, and it's all it can think about"); the factual subject (e.g. "Anthropic's new paper", "Anthropic built a mind-reading tool") must appear in the first half.
• Do not build a title only around a tool name, for example writing only Claude Code 翻车原因, unless the video topic is that tool malfunctioning.
• Do not let the title deviate from the video, for example the topic is Model × Harnessbut the title becomes an ordinary AI 工作流提效。
• Do not abuse "takeover", "armor break", or "zero errors" just to chase trends; use these words only when they fit the topic.
• Do not write all candidate titles into the file by default; keep the file clean first.

Reference

The two-part title pattern and examples accumulated in this workspace are in references/two-part-video-title.md. Read it when you need to generate candidates or explain the commonalities of the benchmarks.