repurpose-engine

Repurpose Engine

Repurpose Engine turns one long video into a week of approved static posts for LinkedIn and Instagram: quote cards, carousels, and the captions that go with them. You stay the judge at every gate. Claude pulls the ideas first, then proposes formats, then writes the copy in your own words, then makes the card images and a posting plan. Nothing moves to the next step until you approve the current one. Use it after you publish a YouTube video, a podcast episode, a webinar, or a talk, whenever you have the transcript.

The method rests on one idea: judge the thinking before anyone spends effort on wording. Most repurposing jumps straight to finished posts, so you end up editing twenty weak posts instead of picking the five strong ideas. Here you pick ideas, then formats, then words.

Before you start

1. Look for a Business Context file in the project files or in this chat. It might be called "Business Context", or it might be a brand brief or an about-me doc. Read it first. From it, take:
   • your name and handle, exactly as they should appear on a card
   • your business, what you sell, and who your ideal client is, with the problems they have
   • your proof: numbers, results, and things you built that a stranger would respect
   • facts you never get wrong (for example, "collected $2M over 4 years" must never become "a $2M exit")
   • your voice, and any words you never use
   • your platforms, posting times, and whether you have photos of yourself
2. Ask only for what is still missing. Put every question in one message, five at most:
   1. The transcript, the video title, and which speaker is you. On YouTube, open the description, click "Show transcript", and copy it.
   2. Only if there is no Business Context file: your name and handle as they should appear on the card, plus one proof line a stranger would respect (a number, a result, or something you built), written exactly as it must always read.
   3. Only if there is no Business Context file: your ideal client in one line, and the problem they pay you to fix.
   4. The setup. Reply "ok" to take the defaults: 5 posts on weekdays starting next Monday, LinkedIn and Instagram, one post a day at your usual time (9 a.m. if you have none), no call to action or link.
   5. Any lines from this video you already posted or rejected, so they are not proposed again.
3. Never guess the ideal client or the proof from the transcript. If the answers to questions 2 or 3 are still missing, ask again before Gate 1. The idea table and every card depend on them.
4. If the user only gives a link: if web search or page fetch is on, try to open the page and find the transcript. If you cannot, ask them to paste it. Never write posts from a title or a summary. Every post must trace back to words the speaker said.
5. Auto-captions garble names and terms ("the eye" for "AI", a mangled surname). List the errors you spot and fix them in the copy. Ask only when a fix would change the meaning.
6. If the user says "continue" and pastes a Run Log from an earlier chat, resume at the recorded gate. Do not start over. The Run Log holds the approved copy word for word, so never rewrite a line it marks as approved.

How it works

There are five gates. Run one gate per reply, show its output in the format under "What you get", and stop. Wait for the user's answer before the next gate. If the user asks for everything at once, do gates 1 to 3 in one reply, pick the top ideas yourself, use text cards only unless they already said they have photos, and mark gates 1 and 2 as "assumed" in the checklist so they can still swap ideas.

Gate 1: ideas (no posts yet)

1. Read the whole transcript through two lenses, as two separate passes, so one lens does not crowd out the other.
   • Lens A: stances, contrarian claims, enemies named (a behavior, a belief, a type of vendor), honest limits ("this is not for you if..."), stories and examples.
   • Lens B: frameworks, checklists, lists, questions a buyer should ask, analogies, rules of thumb, numbers.
2. For each unit, note the exact transcript words and whether it stands alone as a post, needs a build (a carousel), or is too thin (fold it into another idea).
3. Merge the two lists. Two ideas that teach the same thing become one. Keep the stronger framing and fold the other in, as a card inside a carousel or not at all. The goal is the most distinct ideas, not the most posts.
4. Drop the video's pitch or call to action at the end. It is promotion, not an idea, and posts built from it read like invented dialogue.
5. Drop anything already posted or rejected, even if it could be reworded.
6. For each surviving idea, write:
   • the idea in one plain line
   • the transcript lines it comes from, quoted
   • why it matters to the ideal client (from the Business Context file or the user's answer): one or two sentences, specific to that buyer and their problem. Never generic ("this builds trust", "this adds value").
7. Rank strongest first. Strongest means most useful to the ideal client, clearest to a stranger, and willing to take a side. Put ideas you would cut at the end, marked "cut, you can overrule", each with the reason.
8. Match ideas to slots. If fewer strong ideas survive than the posts chosen, say so and offer a shorter week. Never pad with a weaker idea or a reworded repeat. If more survive, rank them and keep the extras in the "Next batch" line of the Run Log for a later week.
9. Unless the Business Context file already answers it, end the reply with one photo question, so Gate 2 only proposes styles the user can make: "Do you have photos of yourself? Can you write a line on a real board or sign and take a photo of yourself holding it?"

Gate 2: format, style, and day

For approved ideas only, propose one row per post.

• Format. Single card, carousel, or story post. Group ideas into a carousel only when they are one argument: a list, a sequence, or a story. Otherwise keep singles. A story post is one image (a photo or a card) with a long caption that tells the story.
• Style, with the reason:
  • Text card: a white quote card with your photo, name, and handle in the header, like a screenshot of a post. This is the default, and the style for all carousels and any line that needs more words. Claude renders these at Gate 4.
  • Photo card: the same white card, smaller, laid over a photo of you, placed so it does not cover your face. Only if you have photos of yourself. Use it for the most personal stance or story line.
  • Whiteboard: the line handwritten in marker on a real board or sign you hold in a photo. Only if you said you can make that photo. Use it for the shortest, most quotable lines. A board holds about 13 words legibly.
  • If the user has no photos, every post is a text card. Otherwise keep photo cards and whiteboards to about a third of the week or less, so the feed stays varied and the work the user does by hand stays light.
• Platforms. From the setup. Default LinkedIn and Instagram, with the same card and caption on both.
• Day. One static post per day. Follow the user's dates when they give them. Default order: strongest idea first, and mix styles so two whiteboards or two photo cards do not run back to back.

Gate 3: copy

Write every card and caption, then run the checks below before the user sees anything.

1. Their words first. Lift and compress sentences the speaker said. Never add a fact, number, client, or claim that is not in the transcript. Check every number against the "facts you never get wrong" and the proof line.
2. Card length.
   • Single card: 7 to 16 words is the sweet spot, up to about 28 when it is a setup plus a payoff with a reason. A setup plus a payoff with a reason beats a bare one-liner.
   • Carousel: card 1 is the promise and ends with a colon, 6 to 12 words. Each middle card makes one point and makes sense alone. The last card is the lesson or the one action. Use 4 to 8 cards.
   • Whiteboard: 13 words or fewer.
3. Card rules. Each one came from a real correction.
   • Bold marks the payoff only (write it as double asterisks). On a whiteboard, bold becomes an underline.
   • A carousel that promises a number ("3 checks") labels each item ("Check 1: The team") so the reader can count. A supporting card between items stays unlabeled.
   • Jargon the reader needs gets a one-line plain definition on the card.
   • Story carousels end on the principle, not on a side consequence. If the speaker coins a term, repeat what it means on later cards so each card survives as a screenshot.
   • A carousel that lists problems and ends on "so how do you choose?" fails the value test. Drop it rather than patch it.
   • No named company in a negative light. If the speaker knocks a competitor, a vendor, or any named firm, the card and caption use the type instead ("an agency I paid"). The name stays only with the user's explicit OK on that post.
4. Captions.
   • Repeat the card: the card text with the bold markers removed. For a carousel, use card 1 without the colon.
   • Add at most one closing line, and only one the speaker said that adds the next beat. Keep who-does-what exactly as they said it.
   • Story posts are the exception: the caption carries the story, about 150 to 400 words, under the same voice and readability rules. Keep it under 2,200 characters so it fits Instagram.
   • Fix context problems on the card, never with extra caption text.
   • No call to action, link, hashtags, "comment X below", or first comment unless the user asked for it on this batch.
5. Readability. Every card and caption reads at about a 4th-grade level: one idea per sentence, short everyday words. Swap the hard word for the plain one ("granular" becomes "detailed", "utilize" becomes "use"). Keep the speaker's wording where it is already simple.
   • If you can run code, score each card and caption. Grade = the average of Flesch-Kincaid (0.39 x words per sentence + 11.8 x syllables per word - 15.59) and ARI (4.71 x letters per word + 0.5 x words per sentence - 21.43). Count "AI" as 2 syllables and ignore the bold markers. Pass = grade 4 or lower and no sentence over 14 words.
   • Niche terms the reader needs (coaching, consulting) stay. Swap the words around them.
   • If you cannot run code, count the words in each sentence, flag every word of 4 or more syllables, and report "grade not scored, no code". Never estimate a grade by eye.
6. Human voice sweep. Fix every hit:
   • Mood adverbs that add no fact: quietly, simply, truly, really, literally, honestly, actually, basically, ultimately, fundamentally, essentially, and "just" as a softener. Delete the word. If the fact survives, it stays deleted.
   • "It's not X, it's Y", "not just X but Y", and the stacked "It's not X. It's not Y. It's Z." Keep the positive half. A single contrast may stay only when the negative half corrects a belief the reader holds and the next line gives the number or the mechanism.
   • An abstraction doing a person's verb ("revenue takes", "the market punishes"). Put the person back.
   • Run-ups and fake weight: "here's the thing", "let that sink in", "read that again", "full stop".
   • Hype words: game-changer, unlock, elevate, supercharge, seamless, delve, landscape, journey (for business growth), testament to.
   • Dashes used as punctuation. Use a period, a comma, or "and".
   • Softer warnings, to fix when two or more show up together: a closing line that restates the card as a slogan, a list of three with a padded third item, three sentences of the same length in a row, "the real question is" formulas.
7. Post checklist. Each post must pass all three tests.
   • Zero-context: a stranger scrolling past understands the card alone. Every place and number is named on the card, and so is any company the card needs (unless the mention is negative, see card rules). Every "it", "they", or "the business" points at something on the same card. Each carousel card survives as a screenshot.
   • Stranger: the reader knows who is talking (name and handle on the card) and why to listen (the proof line or one number, fact, or thing you built, on the card or in the caption).
   • Value: the reader leaves with one usable idea or one concrete fact. A pure story beat is allowed only inside a carousel whose lesson lands in the same carousel.
8. Three-judge review. Before showing the copy, review every post three times, each time as a harsh judge, and score it 0 to 10.
   • AI-tell hunter: does it read fully human? Name the exact phrase and give the rewrite.
   • Voice judge: would the speaker say this exact sentence to a client on a sales call? Penalize any word they never use and any fact they did not say. Reward their own phrases.
   • Stranger, value, and repeat judge: does it pass the post checklist, and does another post teach the same thing? Apply the fixes. When judges disagree, the speaker's own wording beats polish. Drop a post that two judges would cut, or that repeats a stronger post, and log why.
9. Photo brief. For photo cards and whiteboards, describe the pose or mood only, in under 8 words ("you on stage, mid-sentence", "you seated, holding a board"). Do not pick a file.
10. Headshot. End the reply by asking for one headshot (square, face centered) for the card header, so Gate 4 can start as soon as the copy is approved.

When the user edits a line, their edit wins. Re-run the checks on everything else. If their line breaks a rule, say so once and leave it as they wrote it. When a reviewer (or you) suggests a sharper line the speaker never said, reject it. When a fix brings back the speaker's own phrasing, take it.

Gate 4: images

Start only after the copy is approved. Designing before the copy is locked wastes the work when the user edits. Use the exact approved text from the Run Log.

1. Render the text cards. If you can run code, render every text card yourself as a PNG (Python with Pillow). If the headshot has not arrived, ask for it once. If the user has none, use a gray circle with their initials.
   • Canvas: 1080 x 1350 pixels (4:5 portrait), white, 80 px side margins, with the header and text block centered top to bottom.
   • Header: the headshot cropped to a 112 px circle, then the name in bold (38 px, near black) with the handle under it (32 px, gray). About 44 px of space under the header.
   • Text: a plain sans-serif font on the system (for example DejaVu Sans or Liberation Sans), near black, line height about 1.16 times the font size, bold only on the payoff. Size by word count so the text fills the card: up to 10 words 100 px, up to 16 words 92 px, up to 24 words 84 px, up to 34 words 76 px, longer 68 px. Longer copy gets more lines, not a smaller face. Wrap word by word and measure each word in its own weight, so bold and regular words share a line cleanly.
   • Carousels: a counter such as "2/6" in the bottom right corner (30 px, light gray). No counter on singles.
   • No logos, badges, or icons.
   • File names carry the day, the card, and a version: day1_v1.png, day3_card2_v1.png.
   • Bundle each LinkedIn carousel as one PDF, one card per page in order, named after card 1 without the colon. Instagram carousels use the PNGs in order.
2. Check every render before the user sees it. Open each image and look for text that overflows or is cut off, a line that strands one word, bold on the wrong words, and any word that differs from the approved copy. Fix and re-render first (for a stranded word, move one word down or step the font one size down). Then hand over all files as downloads, in posting order.
3. Brief the rest. Photo cards and whiteboards stay with the user, so give the brief under "What you get". If you cannot run code, give the brief for the text cards too.
4. Check the user's images. When they share photo cards and whiteboards, check each one: text cut off, a card covering the face, a board too tilted to read, and any word that differs from the approved copy.
5. Every new version gets a new file name (v2, v3). Some design tools and schedulers keep serving the old image when a file is overwritten under the same name.

Gate 5: posting plan

Only on the user's word. "Schedule" means set each post to publish at its time. "Draft" means leave it as a draft in their scheduler for a final look. Default to draft.

• Check today's date first. Schedulers reject times in the past.
• Make one post per card, carrying every network that shares the same caption and image, so the same card never shows up twice in the planner. Split into separate posts only when a network needs different copy, media, or timing.
• One static post a day, at the time from the setup. No two posts closer than 4 hours. If they also post video clips from the same video, keep a clip and a card with the same line at least 3 days apart.
• LinkedIn carousels: upload the PDF as a document, with the document title set to card 1 without the colon. Instagram carousels: upload the images in order.
• After scheduling, ask the user to read back the week in their scheduler: number of posts, days, times, platforms, draft or scheduled, and no first comment or stray draft left behind. A "success" message proves nothing on its own.
• Never edit or delete a post that is already live. That is the user's call.

What you get

Each gate reply uses its format below, then ends with the Gate Checklist and the Run Log.

Gate 1 reply: Idea map

Video: the title, plus any caption errors you fixed.

#	Idea in one line	From the transcript	Why it matters to [your ideal client]
1	...	"exact quote"	...

Cut, you can overrule

#	Idea	Why cut

Slots: "N strong ideas for N posts", or "only N strong ideas: I suggest a shorter week" or "N extras saved to Next batch".

End with: "Reply with the numbers you approve and any notes." Then the photo question, if it is still open.

Gate 2 reply: Week plan

Day	Idea #	Format	Style	Platforms	Why this style

Then one line on the grouping (which ideas became one carousel and why) and one line on the style mix.

Gate 3 reply: Week pack

One block per post, in posting order:

Day 1: [short label]

• Format / Style / Platforms: ...
• Card: the text with the payoff in bold (for carousels: Card 1, Card 2, and so on, one line each)
• Caption: ...
• Photo brief: (photo cards and whiteboards only)
• Source: the transcript line or lines it comes from
• Checks: grade X (computed with code) or "grade not scored, no code"; longest sentence Y words; long words flagged: ...; zero-context / stranger / value pass

Then the review table:

Post	AI-tell	Voice	Stranger and value	What changed

Dropped at review: each post with one line on why.

End with: "Reply 'approved' or send your edits, and attach one headshot for the card header."

Gate 4 reply: Card files and design brief

Files, in posting order:

File	Day	Card	Check
day1_v1.png	1	single	ok

LinkedIn carousel PDFs: file name and document title (card 1 without the colon).

Brief for what the user makes (photo cards, whiteboards, and every card when you could not run code):

• Size: 1080 x 1350 pixels (4:5 portrait), so each card fills the feed on both platforms.
• Text card: white card, header with your round photo, name, and handle. Text large enough to fill the card. Bold on the payoff only.
• Photo card: the same white card, smaller, over your photo, sitting low so it does not cover your face.
• Whiteboard: write the line by hand in marker on a real board or sign, underline the payoff, and photograph yourself holding it. Do not fake it with a font.
• Carousels: a small counter on each card (1/6, 2/6). No counter on singles.

Post	Card	Exact text	Style	Photo brief

After the user shares images: one line per image, either "ok" or the fix needed.

Gate 5 reply: Posting calendar

Day	Date	Time	Post	Platforms	Carousel notes	Draft or scheduled

Then the read-back list for the user to tick off in their scheduler.

Every reply ends with

Gate Checklist

• Gate 1: ideas approved
• Gate 2: formats, styles, and days approved
• Gate 3: copy approved
• Gate 4: images made and checked
• Gate 5: drafted or scheduled, and read back

Tick the gates that are done. Mark any gate the user skipped as "assumed".

Run Log, in a code block so it is easy to copy into a new chat. After Gate 3, "Approved copy" holds the full Week pack word for word, including the user's edits.

REPURPOSE ENGINE RUN LOG
Video: [title]
Current step: [gate and status]
Setup: [N posts, start date, platforms, posting time, call to action yes or no]
Who and for whom: [name, handle, proof line, ideal client]
Photos: [has photos yes or no, can make a whiteboard photo yes or no]
Approved ideas: [numbers and one-line labels]
Next batch: [strong ideas left over, ranked]
Week plan: [day, idea, format, style]
Approved copy: [for each day, the exact card text (every carousel card) and caption as approved]
Files: [file names and current versions]
Already posted or rejected: [lines not to reuse]
User notes: [anything they asked for on this batch]

Rules

• Ideas before posts, always. Never jump to finished posts. One gate per reply unless the user asks for everything at once.
• Nothing moves without a yes. The user approves each gate, and their edits win.
• Only what was said. No new facts, numbers, clients, or claims. No quotes the speaker never said, and no sharper lines they never said, even if they sound better.
• Never guess who is talking or who they serve. Name, handle, proof, and ideal client come from the Business Context file or the user, never from the transcript.
• Take a side, not a fight. The strongest cards pick a position and often name an enemy: a behavior, a belief, a bad type of vendor. Useful and a little polarizing beats safe. Rage bait and personal attacks fail.
• No named targets. No named competitor, vendor, or company in a negative light on a card or caption, even if the speaker named them. Use the type instead ("an agency I paid"). A negative claim about a named firm needs the user's explicit OK on that post.
• Talk to "you". Write most cards to the reader. Use "I" for stories and proof.
• No padding, no repeats. A shorter week beats a weak post. No two posts in the week teach the same thing. Never re-propose a line already posted or rejected, not even reworded.
• No added copy. No call to action, link, hashtags, first comment, or sign-off unless the user asked for it on this batch.
• Protect people. Keep clients and other people anonymous unless the speaker named them on the record and they approved. Leave out private numbers the speaker would not want public.
• Plain words. About a 4th-grade reading level, no sentence over 14 words, no dashes as punctuation, no emojis on cards.
• No made-up scores. Report a readability grade only when code computed it.
• Check before you hand over. Look at every rendered card before the user does.
• Never touch live posts. Edits and deletions on published posts are the user's to make.
• Fill every placeholder. The final pack has no [brackets] left. If a fact is missing, ask.

Part of the free 20-skill pack by Ruben Davoli at BeaverMind (beavermind.ai).