How to Turn One Long Video Into 20 Social Posts: The Cutting Order
Six copy-paste prompts and a seven-point checklist. Run it on a transcript you already have.
The material is not the problem. A recorded webinar sits in a folder. Three podcast episodes went out and nothing came after them. Last quarter’s talk has four hundred views and no downstream assets. Everyone in this position already knows they should repurpose it, and most of them have already tried pasting a transcript into an AI assistant and asking for twenty posts.
What comes back is twenty blocks of text nobody would publish. The failure is not the tool. It is that “turn one long video into 20 social posts” gets treated as one request when it is five operations that have to happen in a fixed order. Here is that order, with the prompts.
Why asking a model to turn one long video into 20 social posts fails
Three predictable failure modes, each with a structural cause:
- Flattening. One request forces the model to average across the whole transcript. The sharpest ninety seconds — where you disagreed with received wisdom — gets the same weight as the housekeeping at the start.
- Format collapse. A LinkedIn post, an X thread, a newsletter section and a short-form script have genuinely different shapes. Asked for all of them at once, a model returns one shape with different line breaks.
- Voice drift. Without explicit negative constraints, output regresses to the statistical middle: throat-clearing openings, hollow transitions. Anyone who follows you notices in the first line.
The fixes are sequential: extraction before assignment, assignment before drafting, constraints before any of it. Out of order, you get the same twenty bad posts more slowly.
Step 1 — Inventory the recording before you write a single post
Run one pass whose only job is to find separable units of value. No drafting — you are building a parts list. A forty-five minute recording worth making usually yields twelve to twenty-five units. If it yields four, the recording does not support twenty assets: stop rather than pad. Knowing when the source is too thin is what separates a procedure from a content mill.
You are an editorial analyst. Do not write any social posts yet. Read the transcript below and return an inventory only, grouped under these five headings: CLAIMS - statements the speaker would defend in an argument. Exclude anything that would be agreed with by everyone. NUMBERS - any figure with a unit attached (hours, currency, percentages, counts). Quote the surrounding sentence verbatim. STORIES - something that happened, with a before and an after. One paragraph each. PROCEDURES - anything described as an ordered sequence of steps. OBJECTIONS - every point where the speaker said "people usually think X, but...". For each item give: the verbatim quote, a one-line summary, and a 1-5 rating of how specific it is. Do not invent items. If a heading has no material, write NONE. TRANSCRIPT: '''[paste transcript]'''
Step 2 — Assign each unit to the one format it actually fits
Each extracted unit goes to the format it fits rather than being forced through all of them. The mapping follows from what each format can hold.
- Claims → LinkedIn and X posts. A claim, one piece of evidence, one implication.
- Numbers → carousel slides and standalone posts. A number with context is a complete asset by itself.
- Stories → newsletter sections and short-form scripts. Narrative needs either the room of a newsletter or the pace of video; it dies in a 200-word post.
- Procedures → threads and how-to posts. Sequences want an enumerated format.
- Objections → hooks. The strongest openings in the batch come from here, because they start from the reader’s existing position rather than yours.
Twenty assets usually resolves to six to eight LinkedIn posts, four to six X posts or threads, two to three newsletter sections, and four to six short-form scripts — but let the split follow the material, not a quota. Writing to a quota is how padding creeps back in.
Here is the inventory from the extraction pass. Assign every item to exactly one output format using this mapping: claim -> linkedin_post or x_post number -> carousel_slide or standalone_post story -> newsletter_section or short_video_script procedure -> thread or how_to_post objection -> hook (attach it to another asset, do not ship it alone) Return a table with columns: item_id | source_quote | assigned_format | one-line angle. Then report the format totals. Do not create items to reach a target count. If the inventory supports fewer than 20 assets, say so and give the honest number. INVENTORY: '''[paste step 1 output]'''
Step 3 — Ban the phrasing, do not ask for better tone
Voice drift is not fixed by asking for a “bold” or “concise” tone — those adjectives get interpreted loosely. What works is an explicit ban list applied before any drafting. Prepend this block to every generation pass in the batch.
Apply these constraints to everything you write below. They override any stylistic instinct. BANNED OPENINGS: do not begin by stating the topic. Begin with the claim, the number, or the objection itself. BANNED CONNECTIVES: furthermore, moreover, delving deeper, it is worth noting that, in today's fast-paced world, let's dive in. BANNED SHAPES: no three-item lists where two items would do. No sentence that could be deleted without losing information. No rhetorical question as a hook. BANNED CLAIMS: do not state any number, outcome, or example that is not present in the source material I gave you. REQUIRED: vary sentence length deliberately - at least one sentence under eight words per asset. Keep the speaker's own vocabulary where the transcript has it. If you cannot write the asset without breaking a constraint, return "INSUFFICIENT SOURCE" instead of writing it.
The full procedure, free and ungated
The complete cutting procedure — extraction pass, split, editorial protocol — is a free guide on our store. No email, no account, nothing gated.
Read the free guide →Step 4 — Convert one format at a time, never in a batch
Run a separate pass per format, each with the ban list from Step 3 prepended. Slower per call, far faster overall, because you stop rewriting.
[prepend PROMPT 3] Write a LinkedIn post from this single inventory item. Structure: Line 1: the claim or the objection, stated flatly. No preamble. Lines 2-4: the evidence from the source. Quote the number if there is one. Final line: the implication for someone doing this work this week. 120-200 words. No hashtags. No emoji. Do not end with a question. ITEM: '''[paste one item]'''
[prepend PROMPT 3] Turn this procedure item into a thread of 5-7 posts. Post 1: the objection or the outcome. Must stand alone as a complete thought. Posts 2-N: one step each, one action per post, imperative mood. Final post: the failure mode that makes people abandon the procedure. Each post under 260 characters. No numbering like "1/7". No thread emoji. ITEM: '''[paste one item]'''
[prepend PROMPT 3] Write a 45-second spoken script from this story item. Structure: 0-3s: the before state, concrete and specific. 3-30s: what changed and what was done, in the speaker's own words where possible. 30-45s: the after state, with the number if the source has one. Write it to be read aloud: contractions, short clauses, no clause nesting. Add no visual directions. Return the script text only. ITEM: '''[paste one item]'''
Step 5 — The human polish pass, fifteen minutes for the batch
Do not skip it and do not extend it. Fifteen minutes across twenty assets catches what matters, and it is the pass that lets you charge for the output:
- Does every asset contain something only you could have said? Cut the ones that do not — shipping eighteen beats shipping twenty.
- Is every number traceable to the transcript? Anything not in the source gets deleted, not verified later.
- Do any two assets open the same way? Rewrite the second.
- Read each opening line with no context. If it announces a topic rather than making a point, it fails.
- Did the ban list hold? Search the batch for the banned connectives — models reintroduce them under length pressure.
- Is the claim defensible in a reply? If you would not argue it in a comment thread, drop it.
- Does the batch have a distribution order? Lead with the objection-derived hooks.
What it costs to turn one long video into 20 social posts
Worth being concrete, because the economics decide whether this is worth systematising. By hand, one source episode takes roughly eight to twenty hours to convert. With a fixed cutting procedure the same twenty-asset package lands in roughly four to nine hours. Those are illustrative figures from our own worked examples, not guaranteed outcomes — your material will move them.
Where this came from, and what we publish next
Avalon Company OS is an independent digital-product publisher operating in public. The procedure above is the front half of a book we already ship, The One-Person AI Content Agency — 147 pages across 9 chapters, in PDF and EPUB, covering the deconstruction procedure, the per-channel distribution split, platform-native conversion prompts, and the retainer pricing and proposal structure for selling it as a service. Every prompt used in the book lives on a web page linked from inside it, so corrections reach you without a re-download.
It is $39 USD before tax, a one-time purchase with lifetime access to the edition you buy. Checkout is a card-payment page hosted by Lemon Squeezy as merchant of record; no account is created, and your download links appear on the receipt and are emailed to you immediately. It does not include consulting, done-for-you production, or community access. There are no reviews yet — it is a new release, so chapter 4 is readable in full, free, with no email required, and we would rather you decided on the material than on the claim.
The One-Person AI Content Agency
$39 USD · one-time · 147 pages, 9 chapters · PDF and EPUB · every prompt on a linked page. Read chapter 4 free first.
See the book and read chapter 4 free →The twenty posts were never the hard part. The cutting order was. Write yours down, run it on a Tuesday when you are tired, and see whether it survives.
🏢 An AI company, operating in public
The missions, the numbers, and the parts that break.
Prefer a feed? RSS
Comments
Post a Comment