GEMINI SETUP
============

Each Gem needs BOTH of its files. They do different jobs.

TEXT2VIDEO   minimax_h3_fl2va_*.safetensors, or FastVideo FastH3
             Text to video only. No picture anywhere. Use this one if you mostly
             type an idea and want a clip out of it.

FIRSTFRAME2VIDEO   the same models
             One picture, used as the FIRST FRAME. Use this one when you already
             have the opening image and want it to move.

FIRSTLASTFRAME   the same models
             TWO pictures: the frame the clip starts on and the frame it ends
             on. The video is the movement between them. This one ASKS you for
             the duration, because the closing picture has to land on an exact
             second.


WHICH ONE

Most of the time you are doing one of two things: typing an idea, or animating a
picture you already have. There is a single-mode Gem for each, and each one goes
deeper on its own problem

TEXT2VIDEO invents everything, so its rules are about pinning down who the person
is, what they are wearing and what the light is doing, or the model gives you a
different person every seed.

FIRSTFRAME2VIDEO has the opposite problem. The picture already decided all of
that, so its rules are about anchoring it in ONE short sentence and spending
every remaining beat on what moves. Write a wall of "preserve this, preserve
that" and the clip comes back nearly frozen, which is the usual way first-frame
video disappoints.

FIRSTLASTFRAME is the third everyday job: you know both ends of a shot and want
the model to fill the middle. Its rules are about the JOURNEY - one continuous
movement that arrives on the second picture - rather than anchoring once and
moving freely. It is the only one of the three that needs a duration.

STEPS

1. Go to Gemini on the web, open the Gem manager, and create a new Gem.

2. Name it after what it does, for example "H3 Text to Video", so you can tell
   them apart later.

3. Add the KNOWLEDGE (upload) file to the Gem's knowledge or files area. That
   file only. Never upload the INSTRUCTIONS file.

   All three single-mode file pairs carry a version in the name, so the chip in
   the Gem shows which one is loaded, for example "Minimax H3...upload) v5".
   Check it after any swap. Current: Text2Video instructions v7 with knowledge
   v4, FirstFrame2Video instructions v14 with knowledge v5, FirstAndLastFrame
   instructions v5 with knowledge v2.

   WATCH OUT: "Disable Knowledge Citations" is greyed out until a knowledge
   file is attached, and replacing that file switches it back OFF. So tick it
   AFTER the file is in, every time, and then press Update.

4. Open the INSTRUCTIONS (paste) file, select all of it, and paste it into the
   Gem's instructions box, replacing whatever is there. Save.

5. Make sure no image or video tool is switched on for this Gem, and use it from
   an ordinary chat rather than an image mode. The instructions tell it to write
   text and never make media, but no wording beats a switch that says "make
   pictures".


WHY BOTH FILES

The full rule file is 36k to 54k characters, far too big for any instructions
box, so it has to be uploaded. But an uploaded file is SEARCHED rather than read
in full every time, so a rule only reaches the model when your request happens to
match it. The pasted file carries the rules that must never be missed and tells
the Gem to search the big file for the rest.

The pasted files are 8.8k to 24k characters, the FIRSTLASTFRAME one being the
longest because it carries every rule the other two earned plus its own. If
Gemini ever refuses one as too long, say so and it can be condensed: the ChatGPT
versions in the folder next door are the same rules cut to fit 8000, and they
work.


TESTING IT

Run these in a fresh chat. The first is the important one.

1.  make me a video of a woman singing on a rooftop at sunset

    Worded as a request for media on purpose. PASS = a code block of prompt text.
    FAIL = an actual video or picture appears. If it fails, say "text only, give
    me the prompt", and tell me whether it recovers on that second turn.

2.  5 second clip, a barista behind the counter looks up and says "your usual,
    right?" then goes back to the machine

    PASS = one code block with a copy button on it, nothing before or after it.
    For the TEXT2VIDEO Gem, check three more things: no alignment line at the top,
    the prompt opens straight at integrated_multimodal_description, and the
    opening sentence already tells you the person's age, hair, clothes, where
    they are and what the light is doing.

Both of those were run against the TEXT2VIDEO files before they shipped and both
came back clean on every rule that can be checked automatically.


CONVERSATION STARTERS

TEXT2VIDEO                               FIRSTFRAME2VIDEO
Write a video prompt from this idea.     Animate this picture.
Make this into a talking clip.           Make her say a line to camera.
Five seconds, one shot, one line.        Five seconds, small movement only.
A short scene with music under it.       Let the wind move her hair.

FIRSTLASTFRAME
Get me from the first picture to the second.
Four seconds, one movement, no talking.
She starts sitting and ends standing.
Same room, he turns to face the camera.

