Lesson 64.1 gave you a winner: one angle grounded in one quoted fact. Now the temptation is to ask AI for "100 scripts" and call it scale. That produces noise — dozens of wordings with no learning question behind them.

Scale in UGC means something narrower: one core message, deliberately varied, so each version teaches you something.

One message, many surfaces

Keep the product truth fixed. Vary the surface. From one approved message, AI can draft:

• Multiple hooks — same proof, three different first 3 seconds.
• Multiple lengths — a 15-second cut, a 30-second full beat, a 6-second reminder.
• Multiple speaking styles — direct-to-camera, voiceover-demo, reply-to-comment, duet-style reaction.
• Caption variants — one that adds context, one that asks a question, one that carries the source or address.
• Thumbnail and end-card directions — first frame, on-screen title, final card.
• Translation drafts — same script in a second language, flagged for fluent review before publishing.

None of these changes the claim. The moisturizer still proves "pea-size, fragrance-free, calm look over two weeks." The expense app still proves "photo lands in the right column in under 10 seconds." Variation is packaging; truth is cargo.

The practical rule: generate drafts fast, publish only after a human checks that every variant still shows the same proof and obeys the same do-not-say list. A translation that accidentally upgrades "looks calmer" to "cures redness" is not a variant — it is a new, unapproved claim.

The test matrix: fix truth, vary one thing

A test matrix keeps you honest. Rows are the fixed truth; columns are the single dimension you change.

Change one dimension per variant. If you change the hook, the situation, and the CTA all at once and the video performs differently, you have learned nothing. You cannot tell which change mattered.

This is also where shot lists earn their keep. Each variant needs its own shot list — visual action per line — so the creator films the proof the variant promises. AI can draft all nine shot lists in minutes; the human checks that each one still witnesses the claim.

Why 100 scripts lose to 6 hypotheses

A hundred scripts feel productive. They are usually one script reworded ninety-nine times, with three accidental claim changes nobody noticed. No question, no comparison, no log. The next round starts from zero again.

Six deliberate hypotheses beat them because each one asks a question:

1. "Does leading with the trash bin outperform leading with the red cheek?" 2. "Does a 15-second texture demo hold better than the 30-second story?" 3. "Does a comment-reply style earn more saves than direct-to-camera?" 4. "Does showing the ingredient list on screen reduce skeptical comments?" 5. "Does 'patch-test at the counter' outperform 'link in bio' for this audience?" 6. "Does the Spanish draft (reviewed) reach a second audience without new filming?"

Each hypothesis names the dimension, the variants, and the signal you will watch (completion, saves, comments, visits — see Lesson 64.4). Win or lose, you keep the answer.

Generate breadth with AI, then prune like an editor. Ask for twelve hooks, keep three that begin inside genuinely different moments. Ask for four lengths, keep two you can actually film this week.

Practice: the 3×3 matrix

Pick your winning angle from Lesson 64.1 and its single approved benefit. Build three hooks × three demonstrations — nine variants, one benefit.

Step 1 — Lock the benefit sentence. Write it verbatim. Example: "Pea-size, fragrance-free application; redness looked calmer over two weeks under the same light (personal observation, not a medical claim)."

Step 2 — Draft three hooks. Same benefit, three doors:

• H1: red cheek close-up — "My foundation separated every winter afternoon."
• H2: trash bin — "I threw away two moisturizers before this plain one."
• H3: pump texture — "Fragrance was breaking me out — not winter."

Step 3 — Draft three demonstrations. Same benefit, three proofs:

• D1: half-face application, morning light.
• D2: day 1 vs. day 14 same-light pair + texture close-up.
• D3: ingredient list on screen + shop-counter patch-test.

Step 4 — Cross them into nine. Name each variant H1D1 through H3D3. For each, write: hook visual + line, demonstration shots, length (keep all at 25–30 seconds for a clean test), caption direction, thumbnail first-frame note, and the learning question.

Example card:

## H2D1 — Graveyard + Half-Face
- Hook: bin with two tubes → "I threw away two moisturizers..."
- Demo: pea-size pump, half-face apply, window light
- Caption: adds shop address + patch-test note (does not repeat video)
- Thumbnail: hand holding plain tube over bin
- Tests: does a rejection story hook outperform a symptom hook for the same proof?

Add translation or length variants only after the nine are clean — mark them as drafts needing fluent review, not finished scripts.

Finish line

Create UGC-TEST-MATRIX.md with your locked benefit sentence, the nine named variants (hook, demo, shot list, caption, thumbnail, CTA), and one learning question per variant. Keep the CTA identical across all nine unless CTA is your test dimension.

Quick verification: read any two variant cards side by side. Can you name in one sentence how they differ? If not, they are duplicates. Then check every card against the do-not-say list — one changed adjective can break the test.

Common failure mode: varying the claim instead of the creative. "Clears acne" in H1 and "looks calmer" in H3 is not a hook test; it is two different promises. Fix the truth, vary the telling.

Check your understanding

1. What six surfaces can you vary from one message without changing the claim? 2. Why must you vary only one dimension per variant? 3. Why are six hypothesis-led scripts more useful than 100 undifferentiated ones?

Next

You have nine deliberate variants. But who should appear in them — a real creator, a synthetic presenter, or no face at all? Lesson 64.3 draws the trust boundary.