ray-kb

Ray KB (spoken-content redistribution)

Adapt an already verified illustrated source manuscript into a ready-to-record spoken-content package. Adaptation is not abbreviation: do not try to cram every paragraph of the long article into the spoken script; select only one judgment from the source manuscript that is best suited to video, and rearrange the problem, evidence, rhythm, and closing action.

Fixed principles

The source manuscript is the sole source of truth for facts, judgments, and sources. The derivative script must record source_draft and source_sha256, and must not add facts, experiences, or performance numbers on its own; after the source manuscript changes, re-verify first, then revise the spoken script.
Do not write "personally tested," "actually tested," or specific results unless the source manuscript and fact list can prove the user really did it. Keep attribution for the source author's practice; do not rewrite it as Ray's first person.
One spoken script carries only one central judgment and keeps only one closing action.
Writing and revising the source manuscript belongs to ray-writer. If a factual problem in the source manuscript is found during adaptation, stop and hand it back; do not fix facts in place in the spoken script.
Do not publish automatically. Unless the user explicitly requests it, do not generate finished image cards and do not call paid video or image generation.

Input gate

Accept only kind: draft、kind: article or an already published complete body. When receiving content-pack, research, or review cards, follow the draft link in the frontmatter to find the source manuscript; if it cannot be found, stop and explain the gap, and do not treat the task description as an article.

Do not hard-code the knowledge base root directory: prefer the user's explicit path, then search upward for .ray-obsidian.json or a compatible directory structure. If no compatible knowledge base is found and the user only wants a one-off spoken script, the content package may be delivered in the current workspace, with a note that this time it did not enter knowledge回流; if a long-term pipeline is needed, first hand it over to ray-obsidian。

Read the source manuscript, finished package, and fact list, and confirm:

Which are verified facts.
Which are the source author's views.
Which are judgments Ray is willing to sign their name to.
Which personal experiences the user has confirmed may be used.
Which numbers, product states, and limitations need to be re-verified before recording.

Choose the video angle first

Extract at most three candidates from the source manuscript, and for each candidate write clearly: the target audience, one central judgment, the opening conflict, the strongest evidence, the visuals that can actually be shot, and what would be lost if it were deleted.

Selection criteria are applied in order:

The judgment still holds after removing the tool name.
Existing facts are sufficient to support it; no need to invent a story.
A change, contradiction, or counterintuitive judgment can appear in the first 8–10 seconds.
The middle section has a center of gravity that can be proven by screenshot, screen recording, or real operation.
The ending can give a low-friction action.

When the user does not specify a duration, default to a 2–3 minute mixed spoken script with a real person on camera plus screen recording or screenshots; for 60–90 seconds keep only one reversal and one action; for tutorials over 5 minutes, the demonstration steps and spoken explanation must be layered.

Rewrite as spoken script

Use this progression line:

判断或冲突 → 观众熟悉的挫败 → 事实或真实演示 → 再判断 → 一个动作

When writing, follow:

The first sentence gives a judgment, change, or conflict directly; it does not start with background or self-introduction.
One paragraph states only one action; when a sentence exceeds about 45 Chinese characters, split it first.
Key judgments stand alone as sentences. When three terms appear in a row, insert plain-language explanation, analogy, or imagery.
Spoken delivery includes only what the audience must understand. Full links, commands, numeric definitions, and supplementary steps go in the sources and notes section.
Visuals must prove one line of spoken delivery; they cannot be mere decoration. Label every shot as A-roll, screen recording, screenshot, text card, or later B-roll.
Keep only one closing action. Do not simultaneously ask for likes, follows, comments, shares, group joins, and course purchases.

Spoken delivery tone

Speak like someone who has done the work and is willing to give a judgment to a friend: eye level with the audience, no guru act, no reading from a script.
Technical terms must land back on real actions, results, or costs; internet feel comes from scenes the audience knows and colloquial rhythm, not from hot memes, exclamation marks, and emoji.
Emotion comes from the real contradiction of the topic; you may use rhetorical questions, pauses, short sentences, and local acceleration; do not fabricate anger, shock, tears, or on-site dialogue.
After the byline is removed, the audience can still hear that this is Ray's judgment, not a generic AI explanation.

Output files

Derivative drafts adapted from long-form articles go together with the parent draft in <vault>/10-创作/30-文章草稿/, with the filename adding -口播稿; content drafted as spoken delivery with no image-text parent draft goes in <vault>/10-创作/20-口播草稿/, and it is itself the parent draft. A derivative draft is not a second parent draft; the frontmatter must contain at least:

kind: oral-script
status: ready-to-record
source_draft: "[[母稿文件名]]"
source_sha256: "母稿完整文件的 sha256"
content_pack: "[[对应成稿包]]"
target_platform: "视频号 / 抖音 / B站"
target_duration_seconds: 150
video_format: "A-roll + 屏幕录制"
central_judgment: "观众看完只记住的一句话"
image_card_status: not-started

The body must contain:

## 选题卡: audience, existing understanding, central judgment, evidence, main goal, and the single action.
## 三个开场: one each of counterintuitive, change result, and specific scene, with the final chosen item clearly stated.
## 逐字稿: include only text that can be read directly; do not mix in parenthetical cues, tables, or shot descriptions.
## 拍摄节奏: time, spoken-delivery keywords, visuals, subtitle highlights, and material status.
## 素材清单: four statuses: available, to be recorded, to be made, and can be deleted.
## 图片长文接口: record only the visual proposition, suggested page count, and parent draft source; the default status is not-started。
## 来源与录前核对: original links, volatile information, and pre-recording checklist items.

Image long-form interface

This round does not turn image long-form into a complete production process; it keeps only three stable inputs:

A one-sentence visual proposition.
A suggested narrative order of 6–10 pages.
Parent draft path, content fingerprint, and permitted image sources.

When building an independent visual capability later, start from these three inputs; do not guess the judgment again from the long-form article, and do not copy the identifiable layout of a third-party WeChat official account.

Validation and handoff

After completion, run:

python3 scripts/oral_script_check.py <口播稿.md> --source <母稿.md>

If the check fails, revise and rerun. A passing script only proves that the structure, source fingerprint, and estimated duration hold; a human must still confirm that there is only one central judgment, that the visuals can prove the spoken delivery, that there is only one closing action, and read it aloud once in full; when a real read-aloud is not possible, it must be clearly stated that the duration is only a word-count estimate and cannot be passed off as a measured recording duration.

When visuals are needed, hand the spoken-delivery script path, central judgment, shooting rhythm, and material list to ray-broll; it only adds suitable visual metaphors or explainer segments and does not reverse-modify facts or spoken-delivery judgments. Live-action appearance and screen recording are handled by the shooting plan.

Delivery requirements

Deliver to the user the master script path, the spoken content package, the estimated duration, and the pre-recording checklist items, and clearly state whether an actual read-through was completed; briefly explain which angle was chosen and which candidates were abandoned. Do not write the internal review process as a long report.