A week of content in one click — and what to review afterwards

The button that generates a whole week is the easiest feature to demo and the easiest to get wrong. Easy to demo because twenty pieces appearing on screen is impressive. Easy to get wrong because, without a good review state, it just automates the production of bad content — at scale.

What the batch asks

Less than you would expect. In EverFeed, the batch form basically asks how many pieces of each type: how many posts, how many carousels, how many reels, how many standalone captions.

It does not ask for subjects. The subjects come from the brand kit — and that is the difference between a useful batch and twenty variations of the same post.

Where the topics come from

A topic is a proposed subject. EverFeed extracts them from the kit itself, and each proposed topic cites where it came from: from a persona, a registered offer or a differentiator you wrote.

That solves two problems at once.

The first is repetition. A model asked for “ten post ideas for a bakery” produces ten generic bakery ideas, three of which are the same. A model working over a kit with three personas, five offers and four differentiators has twelve different starting points, and variety comes for free.

The second is auditability. When the topic says “this came from the persona ‘mother who buys in the morning’”, you can judge whether it makes sense before spending a generation. And when a bad topic keeps appearing from the same field, the problem is not the model: it is the field, and it is editable.

What happens when you press the button

The batch does not freeze the screen. Each piece becomes a job in a queue, and each job answers immediately with an identifier — generation and publishing are asynchronous by design.

You can close the tab. The work carries on the server, and when you come back the panel finds the state of each piece again. For a batch of twenty pieces with video, that is not a comfort: it is the difference between using the feature and not using it.

Inside each piece, the pipeline runs step by step — copy, visual, video, caption — and each step stores the provider, the model, the exact prompt sent, how many attempts it took and how long it ran.

“Awaiting review” is the most important state in the system

Here is the thesis of this piece.

An AI content system has, fundamentally, two possible designs. In the first, generating and publishing are the same movement: you configure it, it posts, you find out afterwards. In the second, generating produces candidates, and publishing is a separate decision, taken by a person.

The first design is more impressive in a demo and unsustainable in real life. One wrong piece is enough — an out-of-date price, a missing compliance disclaimer, a video with a six-fingered hand — to cost more than all the hours saved.

That is why the default state of a generated piece is awaiting review, and the creatives board filters by state with that filter prominent. The system does not push anything live on its own.

The design consequence is that review has to be fast, otherwise it becomes the bottleneck the automation promised to remove.

How to review twenty pieces in fifteen minutes

The order matters, because each step eliminates pieces before the more expensive step.

1. Read only the copy, in sequence

Ignore the artwork. Go through the captions and the card titles. You are looking for three things: a repeated subject, a promise you cannot keep and a term you do not use. That discards or fixes the problem pieces before you get attached to an image.

2. Check numbers and names

Price, deadline, opening hours, product name, person name. Those are the errors that cost the most and the easiest to miss, because the surrounding text is well written. If the price came from the registered offer, it is right by construction — which is exactly the argument for registering offers with prices.

3. Lock what is good

This is the step almost everyone skips, and it is what makes the following steps safe. A locked step survives any regeneration: you can work on the artwork freely without risking the caption that came out well.

4. Now look at the artwork

With the copy resolved and locked, redo just the visual step on the pieces that need it. Redoing one step does not touch the others, and the previous take is kept — up to ten — so comparing two versions does not cost a new generation.

5. Check compliance per post type

If your sector requires a disclaimer, check that it is on the piece and legible. The kit stores the disclaimer per post type precisely so this does not depend on your memory on a Friday.

6. Only then schedule

Drag it on the calendar, check the faithful preview of each network — what looks good in a square can crop badly in a story — and schedule per channel.

What to do when the whole batch is bad

It happens, and the right reaction is not to regenerate.

If all twenty pieces came out generic, the problem is almost never the model: it is the kit. Generic is what a model produces when constraints are missing. Go back, write the banned terms, paste two good examples and two bad ones, register the offers with prices. Every field in the kit prevents a specific kind of bad piece — and the fix applies to every future batch, not just this one.

If the pieces came out good but repetitive, the problem is a lack of raw material: too few personas, offers, differentiators. The model cannot vary over what does not exist.

If the videos came out bad and everything else is fine, the problem is usually an inflated prompt — and the fix is taking rules out of the video prompt, not adding more.

The rhythm that works

In practice, the shape that holds up is weekly: one generation session, one review session of fifteen to twenty minutes, and the calendar publishes on its own for the rest of the week.

The gain is not generating fast — it is that review becomes a scheduled twenty-minute activity instead of a daily “what do I post today” decision. That daily decision, not the production, is what makes most small accounts stop publishing in their second month.