How it works
One project, eight stages, nothing retyped.
Throughline is not a prompt box with a nice theme. Each stage produces structured data that the next stage consumes, which is why the script knows which claims it is leaning on and the packaging knows what the script actually delivered.
Set up a brand
Niche, audience, awareness stage, voice, the tones you will and will not use, your evidence standard, restricted claims and calls to action. Every project inherits this, and every project can override it.
Without this, generation is generic. With it, the first draft already sounds like you.
Capture and validate ideas
Ideas go in as one line. Validation scores opportunity, novelty, saturation and audience fit, and suggests angles and audience questions.
Those scores are labelled as inference, not measurement, and come with a list of what to verify. Useful and honest beats confident and wrong.
Diagnose the format
Before research, the project is diagnosed: what kind of video this is, what counts as proof, how time-sensitive it is, what class of claim risk it carries, and which visual grammar is native to it.
This is the step that stops a disaster timeline being written like a self-help essay, and it constrains everything downstream.
Research with the receipts
The competitors a viewer would watch instead of yours, ranked on whether they answer the same question, with what each one actually said, pasted in from its own transcript.
Then the audience in its own words: comments, forums and search questions, synthesised into the gaps nobody has filled. It ends in ten ingredients, from the argument to the evidence to the examples to the statistics to the quotes, and every later step is written from those.
Write the brief, then the script
The brief settles the angle, the promise, the emotion and the opening. The script is then drafted part by part from the research itself, and it is refused outright if there is nothing gathered to write from.
Analysis is blunt: retention risks with fixes, repetition, logic gaps, unsupported claims, and the machine tells that make writing read as generated. The humanize pass rewrites voice without touching a single fact.
Plan the edit
Read each part in the voice you chose, tighten it, and the frames are cut from that recording’s own timings, each with a start, an end and a purpose, then an image prompt per frame in your visual style.
Those times are measured rather than estimated: they come from the voiceover’s character-level alignment. Frames tile the timeline without gaps, and the narration is never altered. Only the picture changes.
Package it honestly
Titles and thumbnail concepts are scored against what the script actually delivers, alongside description, chapters, tags, pinned comment and end screen.
Clickbait risk and payoff alignment are scored explicitly, so a thumbnail that promises a reveal the script does not contain is visible before you publish it.
Repurpose properly
Each platform gets its own hook, structure and call to action: a shorts script written word for word, a thread that escalates, a newsletter that stands alone.
Never a truncated script with a link stapled to it.
What runs where
Bring your own AI
Every step runs on your own key, one of OpenAI, Anthropic (Claude) or Google AI Studio, and that provider bills you at their price. Generation is never resold. Each answer is validated against a schema before anything is stored, and a run that comes back malformed is repaired or refused rather than written into your project.
Never
Your content is never sent to analytics, never used to train anything byThroughline, and never passed to a model provider you have not connected yourself. It leaves only when you run a step, only to that provider, and only under the terms you hold with them.
See it on your own topic
The trial runs a whole project end to end. That is the only honest way to tell whether this fits how you work.