An AI Content Repurposing Workflow: One Video, Six Prompts

A long video holds a blog post, a newsletter issue and a handful of short clips, and most of it never leaves YouTube. The fix is not one “repurpose this” prompt but six small ones, run in order on the transcript, with you checking the output between steps. Below is the whole chain with every prompt written out, the order that keeps it honest, and a table of the official length and character limits each output has to fit.

Why one “repurpose this” prompt fails

A transcript is not an article. Spoken explanations circle back, repeat the key line three times, and lean on what the viewer can see on screen. Paste that into a chat window with “turn this into a blog post, a newsletter and ten captions,” and the model has to make every editorial decision at once: which idea leads, which claims are true, which moment stands alone. It makes those decisions at an average level.

Splitting the job gives you a checkpoint after each decision. You fix the transcript, then you approve the list of ideas, then you let the model write. By the time anything reaches a reader, you have already chosen what it says.

I built a YouTube channel past 130,000 subscribers as a one-person operation, and the video was always the product. Everything else had to come out of it without eating the next shoot. Today I run Claude, GPT and Hermes every day, and this chain is how a finished video becomes text. The script side of the same channel lives in the five-prompt script workflow; more processes like it are under AI workflows.

The pipeline at a glance

The step table below is the template. Each step is one prompt, run in the same chat thread so the model can see the previous output.

Two things stand out. Steps 3 and 4 are written from the idea map, not from the transcript. And the only steps you cannot hand off are the checks in the right-hand column.

Step 1: Start from the final cut, not the raw footage

Run this chain after the edit is locked. I cut in Final Cut, and what I learned doing every step alone is that timestamps and quotes only mean something if they come from the version viewers actually watch. A line that got trimmed in the edit should not show up in the newsletter, and a clip timestamp from the raw recording points to the wrong place.

The easiest source is the caption file. YouTube works with SubRip (.srt), SubViewer (.sbv) and WebVTT (.vtt) files, among other formats, and the first two are the ones its help page suggests for beginners. Final Cut Pro can also generate captions from speech, though Apple lists that feature as U.S. English only and limited to Macs with Apple silicon, and it exports captions as SRT among other formats.

Paste the file into a fresh thread with this prompt:

Below is the caption file from a video I have already edited and published. It has timestamps.

Clean it into a readable transcript:
- Start each paragraph with one timestamp in [mm:ss] form, taken from the file.
- Remove filler words, false starts and words I repeated by accident. Keep every opinion, number, name and example.
- Do not add anything I did not say. Do not correct my facts; mark anything that sounds wrong with [CHECK].
- Start a new paragraph whenever I change subject.

Return only the cleaned transcript.

[PASTE CAPTION FILE]

Skim the result against your memory of the video. If the model merged two subjects into one paragraph, split it now. Later steps inherit every mistake from this one.

Step 2: Build an idea map before writing anything

This is the step that separates a repurposed post from a tidied transcript. You ask for structure, not prose.

Using the cleaned transcript above, build an idea map. Do not write any post yet.

1. The one question this video answers, phrased the way a viewer would type it into a search bar.
2. Every distinct point I make, in order. For each: one plain sentence, the timestamp where it starts, and whether I backed it with an example, a number, or only my opinion.
3. The single point I argue hardest for. Quote my own words for it.
4. Stories or specific moments I describe, with timestamps.
5. CLAIMS TO VERIFY: every number, price, date, rule or policy I state, with its timestamp and the kind of official source that would confirm it.
6. What a viewer would still be missing after watching: questions I raised and never answered.

Part 5 is where the text versions earn their keep. You probably said a number from memory on camera. On a page it will be read slowly, quoted and screenshotted. Check each item against the vendor’s own documentation or the relevant help center, fix what changed, and delete what you cannot confirm.

Part 6 is the other half. The questions a video leaves open are often the best sections of the blog post, because nobody watching could get them answered.

Step 3: The blog post

Write a blog post from the idea map, not from the transcript. Reader: [WHO THEY ARE]. Length: about [N] words.

Rules:
- Open with the viewer's question and my answer to it within three sentences.
- One H2 per main point from the map, ordered the way a reader needs them, not the order I said them.
- Use my phrasing where the map quotes me. Everywhere else write plainly, mixing short sentences with longer ones.
- Keep every number exactly as it appears in the map. Leave items from CLAIMS TO VERIFY marked [CHECK] until I clear them.
- Where a point is easier to show than to explain, write [EMBED mm:ss] so I can link that moment of the video.
- Answer the open questions from part 6 only if the transcript contains the answer. If it does not, list them for me after the post.
- No summary at the end. Close with one thing the reader should do next.

The draft will still sound a little like a model wrote it. Read it aloud once and fix every sentence you would not say; the editing pass in this guide covers the patterns worth hunting. Then replace each [EMBED mm:ss] with a link to that timestamp, so the post and the video send readers to each other.

Will Google treat the post as AI spam?

Google does not ban AI-assisted writing. What its spam policies target is “scaled content abuse”: many pages generated mainly to manipulate rankings rather than help people. The examples include using generative AI to produce many pages without adding value for users, and stitching together content from other pages without adding value.

Its guidance on generative AI content says the same from the other side: focus on accuracy, quality and relevance, and hold automatically generated titles and meta descriptions to the same standard as the body text.

A post built from your own video, with your examples, verified claims and links to the exact moments, is on the right side of that line. A raw transcript pasted under a heading is not. Neither is a script that turns every video into a post without anyone reading the result. If you would not send the draft to a friend who asked about the topic, do not publish it.

Step 4: The newsletter issue

The newsletter is not a shorter blog post. Subscribers already chose you, so give them one idea with a person behind it, not a digest.

Write one newsletter issue for my email list, using the idea map.

This is not a summary of the video. Build the email around the point from part 3 of the map:
- Three subject line options. Each says what the reader gets. No clickbait, no all caps.
- Open with one or two sentences in first person, in the voice of the transcript.
- Body: the point, one story or example from part 4, and why it matters to [WHO THEY ARE].
- Exactly one link: the video, at the timestamp where this point starts.
- Under [N] words. No headings. No bullet list unless I made a list in the video.

One link is deliberate. An email that points to the video, the post, a product and two socials asks the reader to choose, and most choose nothing. If you are still setting up the list itself, start with email marketing for creators.

Step 5: Pick the short-form moments

Automatic clipping tools are good at finding energy and weak at finding a finished thought. Ask the model to reason about completeness instead, then do the cutting yourself in your editor.

From the cleaned transcript, find moments that work as standalone vertical clips.

For each candidate give:
- start and end timestamp
- the first and last sentence spoken
- why it stands alone: does it make a complete point without the rest of the video?
- its length in seconds

Rules:
- A clip starts at the beginning of a sentence and ends after a finished thought. Reject any moment that needs earlier context.
- Keep every candidate under [MAX SECONDS] seconds.
- Rank them. No more than five.

Then watch every candidate in the timeline. The transcript cannot tell you whether you were looking at the camera, or whether the key line happened over B-roll that makes no sense in a vertical crop. Keep the two or three that survive. Five mediocre clips do less than two good ones.

Platform limits the outputs must fit

Every repurposing guide says “adapt to each platform” and then leaves the limits out. These are the numbers from each platform’s own help pages and developer documentation, checked on September 28, 2026.

Two notes on the table. Meta’s API reference allows reels up to 15 minutes when published through the API, while the in-app help page says 3; plan for 3 unless you publish through a scheduler that uses the API. And the chapter rules matter here because the description you write for the long video is where your timestamps live, the same ones the blog post and newsletter link to.

Step 6: Captions sized to each platform

Paste the limits into the prompt. Models will not reliably remember them, and they count characters badly, so ask for the count and then confirm it with a character counter before you post.

Write captions for the clips I chose:
[PASTE CHOSEN CLIPS WITH TIMESTAMPS AND FIRST/LAST SENTENCES]

For each clip:
- a YouTube Shorts title under 100 characters
- an Instagram Reels caption under 2,200 characters whose first line makes sense on its own
- a TikTok caption under 2,200 characters
- a Threads post under 500 characters
- an X post under 280 characters

Rules:
- Each caption says what the clip actually shows. Promise nothing the clip does not deliver.
- Use my words from the transcript where they fit.
- No more than three hashtags per caption, and only ones that describe the topic.
- Point to the full video in words, not with a pasted link.
- After each caption, print its character count.

Being under the limit is the floor, not the target. A 2,200-character reel caption is allowed; almost nobody reads one. The first line does the work on every platform, so spend your editing time there.

Running it every week without it running you

The chain only pays if it survives a busy week. Treat it as the last step of publishing a video rather than a separate project: once the upload is live, export the captions, and run steps 1 and 2 the same day while the video is fresh in your head. Steps 3 to 6 can wait for a slower morning.

Skip the chain for a video that is mostly visual, a video whose facts will be stale within a month, or one where the idea map comes back with a single thin point. Not every video deserves four outputs. How this fits next to the rest of a one-person company’s admin is in using Claude for business operations.

Take the last video you published, export its caption file, and run Prompt 1 and Prompt 2 today. If the idea map has at least three solid points, you have your next blog post.