Turn 1 recording into a finished video

/goal Turn the raw recording at the path below into one
finished, polished, upload ready video, edited entirely by you.
Nobody is going to fix your output by hand, so the file you
produce is the file that gets posted.

WHAT YOU ARE GETTING:
A single talking head recording, already trimmed to its final
spoken cut. No graphics, no captions, no motion, no titles.
There is no script, no transcript, no shot list and no
storyboard. You work out what it needs from the picture and the
audio together.

WHAT DONE LOOKS LIKE:
One video file next to the source, named <source
name>-edited.<ext>, that a person could post without touching
it. Same resolution, frame rate, aspect ratio AND duration as
the source, to the frame. Captions on screen and in sync. Motion
graphics on the moments that earn them. If you would not post it
yourself, it is not done.

STEP 1, LOOK AT WHAT YOU HAVE
Probe the source and record duration, resolution, frame rate and
audio sample rate. Print them. Every later decision hangs off
these numbers. If ffmpeg and ffprobe are not on this machine,
install them.

STEP 2, TRANSCRIBE IT, LOCALLY, WITH WORD LEVEL TIMINGS
Use Whisper, running locally. Do not send the audio to a hosted
API.

Work out what this machine already has before you install
anything. Whisper comes in several forms and any of them is
fine, so search for one that is already here, and only install
one if there is nothing. Whichever you use, it has to emit per
WORD timestamps, not just per sentence, and you may need a
specific flag or output mode to get them.

Print which Whisper build, which model size and which flags you
chose, and why, before you run it.

Then, whichever one you land on, these three things are true and
they will silently ruin the edit if you skip them:
  1. VERIFY THE TRANSCRIPT COVERS THE WHOLE FILE. Compare the
     last timestamp against the real duration from step 1. A
     transcript that stops partway through parses perfectly and
     looks completely valid, and every cut you derive from it
     will be wrong.
  2. UNDERSTAND WHAT TIME BASE YOUR TIMINGS ARE IN. Tools that
     run voice activity detection often report offsets with the
     silence squeezed out, which do not map back onto the real
     file. Prove your word times line up with the source before
     you trust them, by spot checking a few words against the
     audio.
  3. TREAT WORD BOUNDARIES AS ESTIMATES. Speech to text
     stretches a word to fill the gap it lands in, so a word
     time is approximate. When a graphic has to hit a word
     exactly, check that word against the waveform rather than
     trusting the number.

STEP 3, DO NOT TOUCH THE TIMING
The cut is already made. The recording you have been given is
the finished spoken cut, and it is locked.

  - Do not remove a single frame. Do not tighten a pause, do not
    trim a stumble, do not shorten the top or the tail.
  - Do not reorder anything, do not change the speed, and do not
    change the pitch.
  - The finished file is the exact same duration as the source,
    to the frame. That is a hard requirement and you should
    check it at the end.

You are editing the picture, not the timeline. Everything you
add sits on top of a cut that does not move.

STEP 4, DECIDE WHAT THE VIEWER SEES
This step is the whole job and I am deliberately not telling you
how to do it.

Read the transcript as content, not as a caption source. Most of
this you can work out from the audio alone. But a few lines
refer to something visible, like where I am pointing, what I am
holding up, or how my hands move. When a line depends on that,
look at the source frames around it before you animate it, and
use enough adjacent frames to see the motion. Work out what this
video is actually about, where it turns, and what a viewer needs
to SEE at each moment to follow it. Then design that. The
vocabulary of graphics is yours to choose.

Look at what else is in the folder beside the source video.
Anything you find in there belongs in the video. Work out from
the spoken line where each one goes, what part of it matters,
and how much of it the viewer needs to see.

Four things I will say, because they are craft rather than
taste:
  - Anchor everything to the exact moment the words are spoken,
    using the word timings. A graphic that lands a beat late is
    worse than no graphic.
  - Many moments need nothing. Restraint is part of this. A
    graphic on every line is worse than a graphic on the right
    lines, and knowing which lines to leave bare is most of the
    skill.
  - Do not describe on screen what is already being said out
    loud. Show the thing, do not label it. A rectangle with a
    sentence in it is the failure mode, and it is what a lazy
    pass always produces.
  - Design it so a viewer with the sound off still gets it.
Print your plan as a table of timecode, spoken line, and what
you chose to show, with your reason. Include the moments you
deliberately left bare, and why. I will judge the video on this
table as much as on the render.

STEP 5, BUILD IT
Every graphic is built as code, in Remotion. Install it fresh
from npm into your own working folder, the normal way. Read the
API docs at the source repo when you need them:
https://github.com/remotion-dev/remotion.git
Do not clone it and do not build Remotion from source.
  - Scaffold a Remotion project, pin the version, and match the
    composition to the source resolution and frame rate from
    step 1.
  - Build every component for this video from scratch, here, in
    this session. Do not reach for a preset, a template, a
    starter project or a workflow you already have sitting
    around. The design choices are the job, so a borrowed one
    defeats the point.
  - Do not invoke an installed skill, plugin, extension or saved
    workflow for any part of this. Build it here, in this
    session, from scratch. If you already have something that
    does this job, ignore it. The design choices are the job.
  - The speaker's footage is the base layer for the whole video.
    It never disappears, and the speaker is never fully covered.
    The 1 exception is where a spoken line explicitly hands you
    the frame. There, and only there, you may cover the speaker
    completely for as long as the line allows.
  - Captions are required. Word level, in sync, legible on a
    phone, and never more than a handful of words on screen at
    once.
  - Build a real component for each idea. Do not build one
    generic card and pass it 12 different strings.
  - Nothing sits still for long. Everything enters and leaves
    deliberately, and nothing is left stranded on screen after
    its moment has passed.
  - Respect a safe area. Keep text and graphics well clear of
    every edge, and clear of where a platform puts its own
    interface.
  - Pick one typeface and one weight ladder and hold it for the
    whole video.
  - REALISM BEATS THE BRAND PALETTE FOR ANYTHING QUOTING A REAL
    THING. If a moment refers to a real surface, object,
    document or place, render it as that thing looks, not as a
    branded diagram of it. A terminal, a file tree, a code diff,
    an editor, an inbox, a spreadsheet, a map, a receipt, a
    form, a chart, a document page: all of them keep their own
    colors, their own layout and their own furniture. Terminal
    green stays green. This rule OVERRIDES the brand block
    wherever the two disagree, and it applies whether or not you
    keep that block.

BRAND, REPLACE THIS BLOCK WITH YOUR OWN OR DELETE IT
Follow this exactly. If you delete it, choose a look that suits
the content and hold it consistently.

  Background: #f7f9fb, white #ffffff surfaces. Light, never
  dark.
  Ink: #191c1e, dimmed #2a2838, muted #55535f
  Hairlines: #c7c4d7, soft #e6e8ea
  FILLS: burnt orange #c2410c, gradient:
  linear-gradient(120deg,#c2410c,#d97757)
  EMPHASIS TEXT: cyan blue #0284c7, gradient:
  linear-gradient(120deg,#38bdf8,#0284c7,#075985)
  Failure state: #D34135, sparingly and only for a real failure
  Typeface: Inter, with a system mono stack for code and
  terminals
  Shadows: barely there. 0 2px 14px rgba(25,28,30,.05)

  Gradients are the default for a fill, not an accent. For
  gradient TEXT, clip the gradient to the
  glyphs and set color to transparent, or the background
  silently paints over the type.

  Never use teal or green as a brand color. Never introduce a
  third accent.

One rule holds whether you keep this block or delete it, and it
is repeated from step 5 on purpose: realism beats the brand
palette for anything quoting a real thing.

STEP 6, AUDIO
Leave the audio exactly as it is. Do not normalize it, do not
compress it, do not add music and do not add sound effects. Pass
the original audio track straight through to the output.

STEP 7, PROVE IT BEFORE YOU CALL IT DONE
A passing build is not evidence. Do all of these and report the
results:
  1. Render actual still frames across the whole video and LOOK
     at them. Nothing overlapping, nothing off the safe area,
     nothing clipped, nothing blank.
  2. Sort every graphic's start and end time and diff adjacent
     pairs. Two graphics rendering over the same seconds is the
     most common defect in this kind of build, and it passes
     every test that does not explicitly compare them to each
     other.
  3. Confirm the output duration matches the source duration to
     the frame, and that the audio is byte for byte the audio
     you were given.
  4. Check the last caption's timestamp against the real
     duration. Captions stopping early means the transcript was
     truncated. Go back to step 2.
  5. Measure the final loudness and state the number.
  6. State the smallest text size you used, as a percentage of
     frame height. Nothing carrying meaning sits below 3% of
     frame height, and nothing sits within 5% of the frame edge.
     If anything fails either floor, fix it and measure again.
Say plainly which of these you actually ran. Do not claim a
result you did not observe.

WHAT I AM NOT GIVING YOU:
No storyboard, no shot list, no direction on which moments
deserve a graphic and no vocabulary of graphics to pick from.
Those choices are the test. Make them, and defend each one in
the table from step 4.

THE VIDEO:
<paste the absolute path to the recording here>