Read firstFurnace core。
This module is responsible for the complete verbatim script. When the user wants a complete script, deliver body text that can be read directly; do not treat an outline as a script, and do not push it back to Start Furnace. If only one paragraph needs changing, change only that paragraph.
Pitfalls Liu Ye hit when rewriting scripts
Liu Ye said: when he took other people's viral hit scripts and rewrote them himself, he basically never got big traffic. The reasons are:
- He couldn't grasp the key points and didn't know what should be said and what shouldn't;
- He would feel that the reason others' data was good was that every sentence made sense and progressed layer by layer, and he was afraid that deleting one sentence would ruin the effect;
- As a result, either he couldn't change it, or after rewriting it was still similar to the original script's structure, making it easy for people to dig it up and call it imitation.
The only rewrite that got big traffic was the airport one (seeCase Libraryitem 3). Its order was reversed:First there is a scene and a purpose, then go find a viewpoint that fits。
Three lessons
- Verbatim scripts are an ammunition depot: when you see content you're interested in, save the verbatim script first; don't rush to rewrite it. The purpose of saving is to find viewpoints later, not to swap synonyms sentence by sentence.
- Match scenes first: when you encounter a good scene, a good hot topic, or a moment you especially want to film, first ask "what purpose do I want to achieve through this scene," then pull out the saved verbatim scripts and see which viewpoint fits best. When there is already a clear theme, you can also start from the theme and search.
- Borrow viewpoints, not structure: it's fine if the viewpoint is the same, but if what the first sentence says, what the second sentence says, and what examples are used are all similar to someone else's, that's copying. First distinguish what viewpoint the reference expressed and what its basis was, then reorganize it using the user's own experiences, demonstrations, and judgments. Sources must be indicated; other people's original words, identities, and achievements cannot become the user's.
Workflow
- Read the task, material, theme, target audience, and the user's speech samples. Distinguish four things: raw facts, the user's own statements, viewpoints in references, and claims that need verification.
- When there is a reference, extract the viewpoints worth borrowing and the progression logic, and mark which ones must be replaced with the user's own material. If material is missing, ask one key question, or clearly mark a placeholder; do not invent experiences or numbers.
- Set the main line and evidence connections. When structure needs help, arrange it according to the methods inStart Furnace, but the final script is still completed by this module.
- Write the complete body text, with every sentence connected. Once the audience feels that consecutive sentences don't connect, they will swipe away. Facts already presented at the beginning (for example, if the beginning already said "what the original price was and what was finally spent"), should not be repeated as-is in the body; the body advances forward (how it was discovered, what the audience should do), not retells. Delete information that does not support the core, and do not omit necessary demonstrations and processes; duration is determined by the user's requirements and the amount of information, and the body text is not cut off for the sake of "one screen."
- Read through it once to check whether the attitude is consistent throughout: if you said earlier "I'm not here to sell you anything," you can't directly solicit customers at the end; if you said earlier "don't jump to conclusions," you can't draw absolute conclusions at the end. If it's inconsistent, change the ending to match the trust established at the beginning, for example, switch to "next time there's another... I'll film it for you to see."
- Review and revise:
- From the perspective of retention, completion rate, and follower growth, go through it cut by cut according to thesix cuts for revising a draft. When Liu Ye revised the AI first draft himself, almost every move was "delete."
- From the perspective of language texture, go through it once according toremoving the AI flavor: first read it aloud and see whether it sounds like this person talking to a friend; then clear out fake contrasts and the ten spoken-delivery items. When the code can be run, first use the
scripts/ai_tell_linter.pyin the main entry directory to scan the draft once; hard blocks should be changed by default. Compare against the user's samples and only change sentences with specific problems. Don't pretend to be the person by adding catchphrases, and don't add inner thoughts out of thin air.
Delivery
Give the main text first. When needed, attach separately: brief shot prompts, sources and information to be verified, and why writing it this way suits this time's audience. When placeholders are included, clearly say this is a draft pending completion and cannot be said to be ready to publish directly. When the user does not want the full draft, do not expand it.
At most one evidence-based revision plus one fact and tone check; leave the remaining disagreements for the user to choose. When the idea is already clear, do not restart the interview.
If the user wants to keep this time's references and materials for next time, read as neededammunition depot and material cards. The generated draft cannot in turn be used as evidence of the user's own experience.