GroundSlate
Prompt recipes
Precise asks that produce finished work.
/
What are you trying to do
Find footage
Cut and assemble

One person, every place, ten minutes of it
One clip a day, years of them, the crop following their face
Vertical that follows the subject, not the middle of the frame
The crop follows a named person's face, not the centre of the frame
A trip, cut to a reel, with rules
State the pacing rules up front, read the plan before anything renders
The whole reel in one ask
One call returns the wired plan, and it renders once you have read itSound and music
Design and render

A software walkthrough that writes its own chapters
Callouts pinned to the pixels you pointed at, and chapters exported
A launch film cut from the site itself
Cut from what the product looks like today, not from March's screenshots
One sales video, forty personalised copies
Build the deck once, render a version per prospect
A deck that presents itself
Send the pitch instead of presenting it
A branded audit, rendered as video
The audit arrives as a film, on brand, rendered on your own Mac
Titles and lower thirds, on brand
Author overlays from a spec, timed to the beats so they land
A grade, and a way out
Bake a LUT from a reference clip, and hand editors real files
A narrator from a sentence
Describe the voice once and every video after it sounds like your channel
Give a character one voice, forever
One identity kit, so every script sounds like the same characterPeople and relationships

A family tree that plays
Every branch gets a chapter of that person's best footage
Teach it who everyone is
Walk the uncertain faces once, and every cut after it gets sharper
See who is on screen and when
Names and timecodes, so shot selection stops being a scrub
Record two relationships, get a whole family
Mark a couple of links and the rest of the family is derivedMap a person's footprint
Every place you have footage of someone, grouped by city
Fix a name and let the words catch up
Rewrite the prose after an identity change, with no model callsOrganise and index
Ten years of loose files, organised overnight
Index the drive, then let it name the events it finds

Teach the vision model your world before it looks
A few sentences of your vocabulary, given to the models before they look

Let it propose what belongs in an event
Ranked candidates from the edges of the envelope, for you to confirmOperate and scope

Watch the queue
See what is running, and have your agent wait for the render
Work from the laptop, without the big drives
Use only what is plugged in, and fail clearly when it is not
Keep client work genuinely separate
Name a profile and your agent works in it, whatever the app has openDrive it from two clients at once
Plan in one client, render from another, one queue behind both

A series
The animated series
A 2D show drawn as SVG and played as HTML. The look is a style kit, every character a puppet, the world as sets, each shot an HTML page GroundSlate captures into the film, and notes a second agent starts from.
Read the five steps

