Forgotten Valley

The Series

A documentary YouTube series on Illinois Valley history. John's voice. Real historical footage + AI-generated assets. Six episodes. Each one bleeds into the next.

Visual through-line: same river, same rock, different people. Every civilization built on the bones of the last one. This is John's actual home — that's the emotional engine.

The "Huh" Moments Are the Whole Point

Not trivia — genuine reframes. Grand Village held 20,000 people and most people drive past it every day. Ottawa hosted the first Lincoln-Douglas debate. The canal that took 12 years to build was obsolete within a decade. These aren't interesting facts — they're the reason someone keeps watching.

Episode 1 — Structural Breakdown

Runtime: 5:23  |  Chapters: 6 + closing  |  VO: John's voice, noise-cleaned  |  YouTube: Published ↗

Structure pattern: Each chapter = one narrative punch. One central "huh" moment. Piano at 0.07 — never competes, never ramps. Every chapter is standalone, then concatted.

How John Makes Visual Decisions

Technical Stack

Environment

Output Spec (all chapters and episodes)

Resolution:   1920×1080
Frame rate:   24fps constant
Video codec:  libx264, preset=fast, crf=18
Pixel format: yuv420p
Audio:        aac 192k 48kHz
Flags:        -movflags +faststart

Ken Burns Pattern (stills)

zoom_end = 1.04 MAX. Never higher. Performance killer above this.

def kb(src, out, seconds, zoom_start=1.0, zoom_end=1.04):
    frames = int(seconds * 24)
    zr = zoom_end - zoom_start
    vf = (f"scale=2400:1350:force_original_aspect_ratio=increase,"
          f"crop=2400:1350,"
          f"zoompan=z='min({zoom_start}+({zr:.4f}/{frames})*on,{zoom_end:.4f})'"
          f":x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)':d={frames}:s=1920x1080:fps=24,"
          f"format=yuv420p")

Video Clips (no zoompan — scale/crop only)

def clip_vid(src, out, ss, t):
    subprocess.run(['ffmpeg','-ss',str(ss),'-i',str(src),'-t',str(t),
        '-vf','scale=1920:1080:force_original_aspect_ratio=increase,crop=1920:1080,format=yuv420p',
        '-r','24','-c:v','libx264','-preset','fast','-crf','18','-an','-y',str(out)],
        capture_output=True, check=True, timeout=60)

Audio Mix

VO:    aac 192k 48kHz, volume=1.0
Piano: 0.07 pre-processed — DO NOT apply additional volume filter
Mix:   amix=inputs=2:duration=longest:normalize=0

Hard Rules

Asset Sourcing Philosophy

50/50 Rule: Half real archival material (LOC, ISHL, Wikimedia), half AI-generated for local gaps where archive material doesn't exist.

Library of Congress — Primary Source

Public domain pre-1928 photos at loc.gov/photos. HABS architectural documentation for canal locks and infrastructure. Brady studio portraits for Lincoln, Douglas, Marquette.

Higgsfield — AI Generation Rules

Visual Language by Era

VO & Timing Workflow

Whisper is ground truth. After John records VO, run Whisper base model → transcript with timestamps → that's the shot list foundation. Not the script file. Not an approximation.

Recording rule: Smart Recorder (Splendio) → clean WAV. NOT the default Android recorder (3GPP/AMR = unusable). Record as ONE continuous take when possible. Dramatic pause needed: two takes joined with silence in ffmpeg.

Chapter Audio File Structure

AUDIO/chapters/chN-vo-final.mp3
Piano overlaid during assembly
VO offset = 5s for Ch1 (title card), 0s for all subsequent chapters
Each chapter is built standalone → concatted in final assembly

The Build Brief as Quality Gate

The brief is the product, not the build. If the brief is wrong, the build is wrong. A locked brief contains:

Sub-agent gets the locked brief. I write the brief. Sub-agent builds. No improvisation on assets or durations without checking.

Episode 2 — "The Settlers" — Current State

What's Locked

Assets On Disk (LOC Archival)

Chapter Build Status

Still Needed

Path Forward — Ep2

The Model: Chapter by Chapter

Ep1 proved this. Each chapter is its own build. Once locked, it's locked. Don't touch it.

Aesthetic Consistency (Ep2 must match Ep1)