Case study
Why your AI writer loses the thread on a long sales script
Short ads from an AI tool can be fine while its long scripts forget the start, repeat a point or come out short. A client's long sales scripts now hold together, and here are the checks to ask for.
Key takeaways
- If a long text comes back with a repeated point or a forgotten promise, the model probably never saw the whole of it in one request.
- Measure every request before it is sent, and have the tool say what it left out rather than dropping it.
- Hold each section of a long text to its own target length, because the model tends to write short.
- Cut long texts at their own paragraph and section breaks, not by counting blocks from the start, so an early edit leaves the rest in place.
- Keep a log of every request and answer; the faults you cannot see on screen appear there first.
The copywriters at a client of ours write long sales video scripts with an AI writing tool we built into the video platform they work in. Their short ads came out well; their long scripts did not. We rebuilt the way the tool handles a long text, it is live on their platform now, and in the first run after the length checks a new script came out within about two per cent of the length asked.
What a broken long script costs the sale
A long sales video works by building: each part sets up the next, so a point repeated or contradicted loses the viewer exactly where the sale is made.
This is what the team saw before we changed anything:
- a section near the end forgetting what an early one had promised;
- the same point arriving twice;
- an edit to one paragraph breaking the bridge to the next;
- a script shorter than the length asked for;
- an answer that stopped halfway with no warning.
A tool that writes a good paragraph but a broken whole costs a team more time to repair than writing the script by hand would have taken.
Why a long script never fits in one request
Every AI model reads a fixed amount of text per request, and the brief, the research, the instructions, the text so far and the answer all share that one space.
A script of the kind that runs twenty minutes or more can fill most of that space on its own. Our tool had treated a long script like a short ad: one request, one answer. Whatever did not fit was lost, and nothing on the screen said so.
What our engineer found
Our engineer measured what every request really carried, using a log added for each one, and the measurements explained each symptom.
- The text of a long script alone filled about a third to about three quarters of the model's room, before the brief, the research and the instructions were added.
- The same ad in Russian, Ukrainian or Arabic needed far more of that room than in English, so the answer was cut at the limit, retried larger, and could still fail.
- The hosting provider's connection dropped any answer that took longer than about a minute and a half to arrive, so the longest answers were lost on the way.
- The old tool cut a script into blocks counted from its start. One shorter paragraph early on moved every block after it, and the team's notes and locks (marks that hold a paragraph as written) drifted onto other words. On the team's own scripts, only about one later block in five stayed in place, and about one in a hundred when the early paragraph became longer.
- The model writes short: five new scripts came out at between about seven tenths and the full length asked.
What we changed
We rebuilt long scripts around the model's real limit, so the whole is planned, written and checked in parts that each know the whole.
- A new script starts with an outline of its sections, which a copywriter approves or edits by hand.
- Each section is written in its own request, carrying the outline, the brief, the research, the sections already written, and a running note of what the viewer knows and feels so far.
- An edit is planned on a map of the paragraphs and made in small batches that each see two paragraphs on either side, with a review of every change before and after.
- A continuity review reads the whole script part by part and judges structure, voice, repetition and every seam, without touching the text.
- Long ads in languages that take more room are written one script per request, and answers now stream, so a long one is never lost on the way.
- Blocks are cut where the script's own words say, so an early edit leaves every later block, its notes and its locks in place. Each section is held to its target length, with one repair.
- Every request is measured before it is sent. If something must give way, the least relevant material goes first and the tool says what it left out; if it still does not fit, the request is refused in plain words. Nothing is cut quietly, and every request and answer is logged.
What the business has now
The new way of writing long scripts is live on the client's platform, tuned by the team's own runs, about three days after the first version of long scripts.
In the first live run after the length checks, a new long script came out within about two per cent of the length asked, where the five runs before ranged from about seven tenths of it to the full length. It took about half as long again to write, because each short section was sent back once to be developed. On all of the team's long scripts, every later block now keeps its words and its notes after an early edit. What the checks still miss in the model's copy is listed for the team, so a copywriter knows what to read for.
What you can check yourself
- Ask your AI tool for a long text and read the last section: does it still answer to what the first one promised?
- Compare the length you asked for with the length you got. If it came out short, ask whether anything holds each section to its target.
- Ask whoever looks after your tool what happens when a request is too long: is it cut, refused, or does the tool say what it left out?
- Ask for the same long text in another language. An answer that stops halfway is the sign of a limit being reached.
- Edit one paragraph near the start of a long text, then see whether any notes on later paragraphs are still on the paragraphs you left them on.
What to do next
Plan a long text as an outline a person approves, then have it written one section at a time, each section carrying the whole in view and a short note of what the reader knows so far. Ask two things of whoever builds or buys your tool: that every request is measured before it is sent, and that the tool says plainly what it left out. Reading the finished text once more for its seams and repeats is still the copywriter's job, and the log of requests and answers is where hidden faults appear first. This is the kind of work our engineers do on the AI tools small businesses rely on.
Questions people ask
Why does my AI writer forget what the start of a long script said?
An AI model can only read a fixed amount of text in one request, and a long script plus your brief and instructions can go past that limit. Whatever does not fit never reaches the model, so the later parts are written without the earlier ones in view. Writing the script in sections, each sent with the outline and a short note of what has been promised so far, is what keeps the whole together.
Why does the AI come back with a shorter script than I asked for?
In our experience a model tends to stop short of a length it is given, and nothing warns you when it does. Ask for each section to be checked against its own target length and repaired when it comes out short or long, rather than checking only the finished text. Then compare the length you asked for with the one you got.
Why does an AI answer stop halfway in Arabic or Russian but not in English?
The same text in Russian, Ukrainian or Arabic takes more of the model's room than it does in English, so the answer reaches the limit sooner. A connection can also drop an answer that takes too long to arrive. Sending one long text per request and streaming the answer back as it is written avoids both.