# The Restructure Prompt **From Chapter 6.2 of *The 4-Hour Side Hustle*.** Turns a spoken transcript into an article draft — without turning your voice into everyone else's. Upload this file as its own Claude Project, or add it to your main Project from Chapter 3.3. It's module 4 of the article chain. --- ## The prompt > Below is a transcript of me speaking about a topic in my field. Turn it into a first draft of a written article. > > **Your job is to restructure, not to rewrite.** > > - Preserve my specific examples, numbers, names, and dates **verbatim**. Do not smooth, generalize, or paraphrase them. > - Preserve my opinions, including ones stated bluntly. Do not hedge them. > - Remove only: filler, false starts, repetition, and things I clearly abandoned mid-sentence. > - Reorder for a reader rather than a listener. Spoken tangents that work aloud usually need to move or go. > - Add subheadings that describe what the section actually says, not generic labels. > > **Hard rules:** > - Invent nothing. Add no statistics, studies, examples, or claims not present in the transcript. > - If a passage is unclear or a claim seems incomplete, do not fix it — mark it `[NEEDS AUTHOR INPUT: what's unclear]` and move on. > - Do not add an introduction that summarizes what the article will cover. Start with the substance. > - Do not add a conclusion that restates the article. > > Target 1,000–1,400 words. Transcript follows. --- ## Why "restructure, not rewrite" is the whole instruction Ask a model to "turn this transcript into an article" and it will produce something readable, competent, and stripped of everything that made it worth publishing. Your blunt opinion becomes a balanced consideration. Your specific number becomes "significant." Your example about a Tuesday in March becomes "in many cases." That's not the model failing. That's the model doing exactly what "write an article" means to something trained on a million articles — regress toward the mean. The four preservation rules exist to fight that, one mechanism at a time: - **Verbatim examples and numbers** — because specificity is the only thing separating your piece from the other four thousand on the topic. - **Unhedged opinions** — because a hedged opinion is not an opinion, and readers can tell. - **Remove only filler** — a whitelist, not a licence. Anything not on the list stays. - **Reorder, don't rewrite** — spoken structure genuinely doesn't work on the page, and that's the one change worth making. --- ## `[NEEDS AUTHOR INPUT]` — the instruction to keep When you speak for twenty minutes you will leave something half-said. A model without an escape hatch will fill that gap smoothly and invisibly, and you will publish a claim you never made and can't support. The bracket makes the gap loud. You'll find them in ten seconds with Ctrl+F, and each one is a place where you actually have something to say and didn't finish saying it. **Never publish a draft still containing one.** In the article chain, route drafts containing `[NEEDS AUTHOR INPUT` to the Needs Edit view and stop them there — that's what the view is for. --- ## The two "do not add" rules **No summarizing introduction.** "In this article, we'll explore…" is a paragraph that tells the reader they haven't started yet. Start with the substance and the reader is already in. **No restating conclusion.** A conclusion that repeats the article is the model's default ending and adds nothing. If the piece needs a landing, you'll write it yourself in ninety seconds and it'll be better. Both rules exist because these are the two places a model reliably pads, and padding is what makes AI-drafted writing recognizable as AI-drafted writing. --- ## Working with the output This is a **first draft**, and treating it as anything else is how the chain starts producing content nobody wants to read. **Fifteen minutes, three passes:** 1. **Search for `[NEEDS AUTHOR INPUT`.** Answer each one. This is the highest-value editing you'll do. 2. **Check every number and name against the transcript.** Not because the model usually gets them wrong, but because when it does, it does so fluently and you won't catch it by reading normally. 3. **Read the opening line aloud.** If it sounds like an article, rewrite it. If it sounds like you starting a sentence, keep it. Then publish. --- ## If output keeps coming back generic The transcript is thin, not the prompt. A transcript where you spoke in generalities produces an article of generalities — the model can't add specificity it wasn't given, and the `invent nothing` rule means it correctly won't try. Re-record with one instruction to yourself: name a real week, a real number, a real person's objection. --- ## Checklist - [ ] Uploaded to a Project alongside the Business Brief and writing samples - [ ] Transcript is you speaking, not written notes - [ ] Target length set to 1,000–1,400 words - [ ] Output searched for `[NEEDS AUTHOR INPUT` before anything else - [ ] Every `[NEEDS AUTHOR INPUT` answered, none published - [ ] Numbers, names and dates checked against the transcript - [ ] Opening line read aloud and rewritten if it sounds like an article - [ ] Chain routes flagged drafts to Needs Edit and stops them there - [ ] Generic output diagnosed as a thin transcript, not a weak prompt