Turn one idea into a week of content

What actually counts as one idea worth a week?

A topic is not an idea. "AI agents for lead generation" is a topic. You can write about it forever and say nothing, which is why so many people do.

An idea is smaller and harder. My test has three parts, and it has to pass all three.

1. 1A specific thing that happened. A recording, a build, a call, a bug. With a date and a place attached, not a general observation about the industry.
2. 2A claim you would defend. One sentence you would still say out loud if a smart person in the room disagreed with you.
3. 3At least one place where you were wrong or surprised. If nothing in the story corrected you, you are reporting, not thinking.

Most weeks I take this from a LadderShift Podcast episode or from something I just shipped. Both pass the test easily, because a real conversation and a real build both contain the parts by default. A blog-style topic you picked off a keyword tool does not.

Why "repurpose this into a week of content" gives you filler

Paste the transcript into a model, ask for a week of content, and you get seven posts in about twelve seconds. They look like this.

• "AI is changing how small teams operate. Here are 5 lessons I learned this week."
• "Most people use AI wrong. The difference is in the workflow, not the tool."
• "Stop chasing tools. Start chasing outcomes."

Fine. None of it tells me what to post on Tuesday. Every one of those could have been written before the episode existed, by someone who was not in it. They are interchangeable, they are the average of everything ever said about the subject, and they have no memory of what actually happened in the room. That is what filler is. Not bad grammar, just no evidence.

The other failure mode is worse and more common: take one paragraph, crop it to three aspect ratios, and post it three times. That is the thing that makes a feed exhausting. It is also the thing Instagram says it filters for. In its creator guidance on original content, it says accounts that primarily repost others' work are not eligible for recommendations, and that changes like adding a border or a watermark or a translation do not make something original.

The extraction step: pull one idea into six parts

This is the whole method, and it happens before you write a single post. You are not deciding what to publish yet. You are taking the idea apart on the table and labelling the pieces.

Six rows. Every row keeps the source attached, which means an exact quote, a timestamp, a filename, or a line from the log. A row with no source is a row you made up, and you will not be able to tell the difference in three days.

If two of those rows come back empty, stop. You do not have a week. You have one post, and forcing it into seven is how the filler gets in.

What the table looks like when it is filled in

Say the episode is with an operations lead at a small logistics company who built an agent to handle supplier emails.

Look at the difference between the mechanism row and the claim row. They are both true, they came from the same hour, and they want completely different formats. The mechanism needs a screen. The claim needs a sentence and an argument. That is the reason one idea can carry a week without repeating itself, and it is the reason chopping the recording into clips does not work: clips split the idea by time, not by kind.

If the mechanism row is the part you keep wanting to build rather than write about, that is a different project. I have written up the smaller version of it in Set up your first AI workflow.

What each platform actually says it wants

I want to be careful here, because most advice about what platforms reward is folklore repeated until it sounds like a fact. So here is only what the platforms publish themselves, and then my reading marked as mine.

Instagram publishes its ranking signals for Reels. It names the predictions it tries to make: how likely you are to reshare a reel, watch it all the way through, like it, and go to the audio page. Reshare is listed first.

YouTube publishes its list of recommendation signals: watch history, search history, subscriptions, likes, dislikes, "not interested" feedback, and satisfaction surveys. Its own help page says the surveys exist to understand satisfaction, not just watch time. Its search and discovery guidance is blunter than most creator advice: focus on what your audience likes, and if people watch, the recommendations follow.

LinkedIn does not publish a ranking explainer of that kind, so I am not going to invent one. What it does publish is format limits, and those are useful on their own. Its video specification lists a maximum duration of 10 minutes and a minimum of 3 seconds. Its document posts accept PDF, PPT, and DOC files up to 100MB and 300 pages, and it says plainly that videos and animations inside a document are not supported and will show as static images.

My reading, and it is only a reading: a platform that asks "would you reshare this" wants a thing that survives being sent to one person with no context. A platform that runs satisfaction surveys wants a thing that was worth the time, which is a longer test than a strong opening. And a platform whose best native format is a flat, silent, 300-page document is telling you it expects an argument, not a performance.

How do you turn one idea into a week of content across three platforms?

Map parts to formats. Not paragraphs to aspect ratios. Here is the week I would actually run from the table above, and the only column that really matters is the last one.

On the Wednesday chapters: YouTube's own requirements are specific, so it is worth getting right. The first timestamp has to start at 00:00, you need at least three timestamps in ascending order, and each chapter has to be at least 10 seconds long. Miss one of those and the chapters simply do not appear.

One more format note. YouTube now allows Shorts up to three minutes, and classifies by aspect ratio: square or vertical is treated as a Short, and a wider ratio such as 16:9 keeps a long video out of that bucket. Worth knowing before you export Friday's clip vertically out of habit.

Where does AI genuinely help in this?

In the boring middle. That is not a small area, and it is most of the clock time.

• Transcribing the recording and timestamping it, so every row in the table can carry a real source.
• Pulling candidate rows out of a long transcript. It is good at finding the four places someone said something claim-shaped in ninety minutes of talking.
• Drafting fifteen versions of one opening line so you can hear which one sounds like you and which one sounds like everyone.
• Checking the week's drafts against each other for repetition. Ask it where you used the same phrase twice and it will tell you honestly, which a human editor is often too polite to do.

Where does it ruin the piece?

At the exact detail that makes the thing worth reading. Ask a model to fill the moment row and it will write that the client was thrilled with the results. Nobody was thrilled. She deleted the agent twice.

So here is the line I hold, and it is a single rule rather than a philosophy: AI may move, sort, compress, and rephrase anything that already exists in the input. It may not produce anything that does not. If the transcript has no number, the post has no number. If nobody objected, the objection row says MISSING and I go and find someone who will object.

This is the same rule I use when an agent touches anything that leaves my hands and goes to a real person, which is most of the work in finding leads with AI. Drafting is safe. Asserting is not.

The extraction prompt, ready to copy

Paste your transcript or your notes into the input placeholder. The rules at the bottom are the part that matters, so do not trim them.

You are helping me pull one idea apart before I write anything.
Do not write posts. Do not give advice.

SOURCE: <episode title, build name, or call date>
INPUT:
<<<
<paste the transcript, notes, or commit log here>
>>>

Return ONE markdown table with these columns:
Part | What it says | Exact quote or line from the input | Where it is (timestamp / file / line)

Fill exactly one row for each of these parts, in this order:
1. The moment. The specific thing that happened, as a scene.
2. The claim. One sentence someone would defend in an argument.
3. The objection. The strongest disagreement, in the objector's own words.
4. The mechanism. How the thing works, step by step.
5. The number or artifact. A count, a file, a screenshot, a table.
6. The thing I got wrong. Where I was corrected or surprised.

RULES
- Every row must carry a real quote or line copied from the input.
- Do not paraphrase in the quote column. Copy the words.
- If a part is not in the input, write MISSING in every column of that row.
- Do not invent numbers, names, outcomes, or dates.
- Do not add a conclusion, a lesson, or a call to action.

After the table, list in plain text which rows came back MISSING and what question I would have to ask to fill them.

The MISSING instruction is doing more work than the rest of it. A model asked for six rows will happily produce six rows, and the fabricated ones look exactly like the real ones. Giving it permission to return an empty row is what makes the output inspectable.

What if the week does not land?

Some weeks it will not. The right question at the end of a week is not "did it perform", because that answer arrives mixed up with timing, luck, and whatever else was in the feed that day.

The question I would ask instead: which row did people actually respond to? Not which post. Which part of the idea. If every reply is about the mechanism and nobody argued with the claim, that is information about what your audience wants from you, and it is not information you could have got by posting the same paragraph three times.

There is more than one honest reading of a quiet week. Maybe the idea was thin. Maybe the formats were wrong. Maybe the claim was correct but so obvious that nobody felt the need to say so. I do not think you can separate those from one week of data, and I would not try.

Why the table is the thing you keep

Posts expire. The table does not.

By the fifth idea you have five of them in one folder, and the rows start referencing each other. The objection from episode two is the claim in episode five. The mechanism you documented in March is the thing you stop having to re-explain. Two ideas that were each too thin for a week become one week between them. That is the compounding part, and it has nothing to do with any platform.

It also survives the platforms changing, which they will. A format limit is a fact with an expiry date. A filled-in table of what actually happened and what you concluded from it is not.

So take the last thing you genuinely finished, the build or the call or the episode, and fill the six rows before you write anything. If four rows are full, you have your week. If two are empty, you have one good post and a question worth asking someone. Both of those are better than seven confident posts about nothing.