Lesson 62.1 locked the brief. Now you need a pipeline that carries the brief to a finished piece without losing it. AI media work is a chain of handoffs, and every handoff is where quality leaks: the script says one thing, the shot list shows another, the voice reads a third.

The nine-stage pipeline

Memorize the order. Loops are allowed, skips are not:

brief → script → system/assets → voice → shots → edit
  → captions/links → review → exports

Each stage has one job:

1. Brief (Lesson 62.1) decides what is true and what success looks like. 2. Script turns the message and evidence into spoken words with timestamps. 3. System/assets locks the visual identity — palette, fonts, character reference, logo placement — before any frame is generated. 4. Voice records or generates the narration from the locked script. 5. Shots generates or captures each visual beat to match the narration. 6. Edit assembles voice, shots, and music into a timed timeline. 7. Captions/links adds burned-in or sidecar captions, alt text, and the verbatim CTA link. 8. Review runs the four human passes from Lesson 62.3. 9. Exports renders masters and placement variants from Lesson 62.4.

Two loops are normal and healthy. Weak-shot → revise-script is the most common: a shot comes back unusable (mangled hands, wrong diagram, off-brand colors) and the cheapest fix is simplifying the script line, not re-rolling the same prompt twenty times. Review → re-edit is the second: the review pass finds a continuity break or a caption error and sends the timeline back one stage, not back to the brief.

What is never allowed is silent skipping — generating shots before the system is locked, or exporting before review. Each skip creates rework that costs more than the stage it skipped.

Six handoff files

A handoff is a file, not a vibe. Each one names its owner and its approval gate: who wrote it, and who must say yes before the next stage starts. For solo builders, owner and approver are both you — but the gate still exists, separated by a break, because fresh eyes catch what tired eyes miss.

| Handoff | Contains | Owner → Approver | Gate |

|---|---|---|---|

| SCRIPT | Lines with timestamps, evidence tags, verbatim CTA, word count | Writer → you | Every factual line traces to brief evidence |

| SHOT-LIST | Shot number, timestamp, visual description, prompt, asset refs | Director → you | Each shot maps to one script line |

| VISUAL-SYSTEM | Palette hexes, fonts, character refs, logo rules, banned styles | Designer → you | New still matches palette + refs |

| VOICE-PLAN | Voice ID, pace, pronunciation, file naming, retake rules | Audio → you | 10-second sample approved before full read |

| EDIT-DECISIONS | Timeline, cuts, music level, lower-thirds, CTA card timing | Editor → you | Timeline matches script timestamps |

| PUBLISHING-CHECKLIST | Files, titles, captions, links, UTMs, alt text (preview of 62.4) | Publisher → you | Every link opens; captions verified |

Write each handoff so a stranger — human or AI — can continue without asking you questions. That means prompts are pasted in full, file paths are exact, and decisions record what was rejected, not just what was chosen. "Voice: calm female, 0.95x, 'Sonariq' pronounced SOH-nah-reek, retake any line with mouth clicks" beats "use a nice voice."

The folder another AI can pick up

Context management from Class 9 applies to media: a fresh session with a clean folder outperforms a long session with scattered files. Use this structure for the class project — a 30-second explainer:

explainer-30s/
  CAMPAIGN-BRIEF.md
  SCRIPT.md
  SHOT-LIST.md
  VISUAL-SYSTEM.md
  VOICE-PLAN.md
  EDIT-DECISIONS.md
  assets/
    refs/          # character refs, logo, palette swatches
    stills/        # generated images, numbered per shot
    audio/         # narration takes + music bed
  working/
    prompts.md     # every generation prompt, pasted verbatim
  masters/
    explainer-30s-master.mp4
  exports/
    explainer-30s-shorts.mp4
    explainer-30s-square.mp4
    thumbnail.png
  PUBLISHING-CHECKLIST.md
  MEDIA-REVIEW.md

The test is simple: open a brand-new AI session, point it at the folder, and ask it to "list the next unapproved handoff and what it needs." If it can answer from files alone, your handoffs work. If it needs you to explain what lives where, the folder failed — rename and reorganize before generating more.

Keep working/prompts.md append-only. Every image, voice, and video prompt goes in with its date and output filename. When a shot looks wrong three generations later, the prompt log tells you what drifted. Without it, you re-roll blindly and burn credits.

Approvals are cheap; rework is expensive

For each gate, define "approved" in one sentence before the work starts. Examples:

• Script approved means: read aloud at pace, lands within 28–32 seconds, every claim tagged to brief evidence.
• Shot-list approved means: every shot has a script timestamp and a system-compliant prompt.
• Voice approved means: sample matches pace and pronunciation; no background hiss on headphones.
• Edit approved means: narration, captions, and CTA card all agree on the same words.

A gate is a pause, not a ceremony. Check the sentence, mark APPROVED — date at the top of the handoff file, move on. Skipped gates do not save time; they move the discovery of the error to the most expensive stage — after export, in public.

Exercise: build the folder

Create the explainer-30s/ folder above for your Lesson 62.1 brief. Fill three handoffs minimum:

• SCRIPT.md with timestamped lines totaling 70–80 spoken words (about 30 seconds), CTA verbatim, evidence tags.
• SHOT-LIST.md with 4–6 shots, each mapped to a script timestamp, each with a full image/video prompt that names the palette and style from the brief.
• VISUAL-SYSTEM.md with palette hexes, font choice, and one character or object reference described precisely enough to regenerate.

Finish line: a folder where a new AI session can read the brief plus three handoffs and generate shot 1 without asking you anything.

Verify fast: paste only the folder contents into a fresh session and ask for "shot 1, ready to generate." If it asks for the palette, voice, or CTA, the handoff is incomplete. Common failure: shot prompts that say "same style as before." There is no "before" for a fresh session — paste the actual hex codes and descriptors every time.

Check your understanding

1. Name the nine pipeline stages in order. 2. Which two loops are normal, and which direction must they never skip? 3. What must every handoff file name beyond its content? 4. Why is working/prompts.md append-only?

Next

The pipeline moves fast. Too fast to trust blindly. Next, Lesson 62.3 adds the human review system — four passes, accessibility rules, evidence checks, and a scored matrix that tells you exactly which two areas to revise.