PROD. SECOND UNIT UGC/SIDES — MODULE 2/DRAFT/Fact-check before publish

Module 2  ·  section 1  ·  1,295 words · ~11 min read

Writing the brief: hooks and scripts that don't sound generated

docDoc type
draftStatus
RequiredFact-check
neverLast verified

1A prompt is a brief, not a wish

The single most common beginner mistake is asking for the thing you want instead of describing the job.

"Write me a hook for a pan." is a wish. You will get back something that could be about any pan, for anyone, because you described nothing that would rule anything out. What comes back will be fluent, confident, and useless — and it will be useless in a specific way that we'll come back to.

"Write three opening lines for a 22-second testimonial, maximum nine words each, no question openers, one concrete sensory detail, for renters aged 24–34 who already own a pan they resent" is a brief. Every clause in it eliminates something.

If you have ever written a brief for a photographer, or received one from a brand, this is the same document. You are not talking to a genie. You are briefing a very fast, very literal freelancer who has never met you, has no taste of their own, and will not ask a follow-up question unless you invite one.

Explain-on-first-use: when this module says the brief, it means the whole instruction you send — the request plus every constraint attached to it. Some people call this "the prompt." Same thing. We use brief because it puts you in the right frame: you are the one commissioning.

2The four things a hook brief has to contain

Everything else is optional. These four are not.

  1. A length constraint, in the unit you'll actually cut to. Not "short." Nine words. Twenty-two seconds. Two lines. Models are poor at "short" and excellent at "nine words," because one is a judgement and the other is a count.
  2. A named audience in a situation. Not "young women." A renter in their late twenties, standing in a kitchen they can't renovate, deciding whether to cook or give up and order in. The situation is what makes the line specific; the demographic on its own does almost nothing.
  3. A vocabulary constraint — what it may not say. This is the highest-value line in any brief, and it is the one beginners leave out. More on this below.
  4. A quantity. Ask for ten. You are not looking for the line, you are looking for a shortlist to choose from, and choosing is your job.

Miss any one of these and you can usually tell which from the output. Missing length gives you something that won't fit. Missing audience gives you something generic. Missing quantity gives you one confident wrong answer. Missing the vocabulary constraint gives you the thing everybody can hear.

3The tell, and why it happens

Watch enough AI-written short-form and you start to hear the same handful of constructions:

  • Stacked adjectives. Rich, creamy, indulgent. Three where one would do.
  • The tricolon. Any list of exactly three, rhythmically balanced. It is a genuinely good rhetorical device, which is why it gets used until it is audible.
  • "Not just X — it's Y." Once a useful reversal. Now a fingerprint.
  • Question openers. Ever wondered why...? Tired of...? These test badly and they are the first thing an unconstrained model reaches for.
  • Filler intensifiers. Honestly. Literally. Actually. Obsessed. Game-changer.
  • Symmetry that nobody speaks in. Real speech is lopsided. Generated speech balances.

None of these are errors. That is the problem. They're the most statistically ordinary way to phrase a thing, and a model with no instruction to the contrary will land there every time, because ordinary is what it is built to produce.

So the fix is not "write better prompts" in some vague sense. The fix is to name the ordinary phrasings and forbid them. That is what the off-limits list in your Brand Voice & Guardrails Intake (Module 3, Form BV-01) is actually for. It looks like a compliance document. It is really a de-genericiser.

The model isn't trying to sound like AI. It's trying to sound like everyone. Those turn out to be the same thing.

4Generate ten, keep one

The ratio matters more than the prompt.

Ask for one hook and you will take it, because you have nothing to compare it to and it will read plausibly. Ask for ten and something useful happens: you stop reading them as answers and start reading them as options, which is the correct relationship. You will typically find one you'd shoot, two you could fix, and seven that are the tell in various costumes.

Seven out of ten going in the bin is not the tool failing. A casting session works the same way and nobody calls it a waste of an afternoon.

Two habits that make the shortlist better:

  • Read them out loud. Every one. The tell is an audio problem more than a text problem — a line can look fine and be unsayable. If you stumble, the camera will too.
  • Cut the first line of any hook you keep. Not always right, but right often enough to be worth trying. Models warm up. The second sentence is frequently where the actual hook was hiding.

5Rewriting is where the real work is

The first output is a draft, and you should treat "make it better" as a non-instruction. Models cannot act on it. Give them the same specificity you'd give an editor:

Instead of Say
"Make it punchier" "Cut it to seven words. Lead with the object, not with me."
"Make it less AI" "Remove every adjective. Keep one concrete noun per line."
"Make it more natural" "Rewrite as something you'd say mid-sentence to a friend, starting in the middle of the thought."
"I don't like it" "Line 2 is the only one with a real detail in it. Give me five more like line 2 and nothing like lines 1 and 3."

That last row is the one to internalise. Pointing at the one thing that worked and asking for more of it is the highest-leverage move in the entire workflow, and it takes about four seconds.

6The line that stays yours

There is a decision inside every good short-form piece that no brief can contain, and it usually looks like nothing on the page.

In the worked example that runs through this guide, the script has a beat in it — hold two seconds on the empty pan, no music — that no model produced. It came from someone watching a rough cut and noticing that the line landed harder when nothing followed it. That's not a prompting failure. That's the job that stayed yours when the other three moved.

You will make roughly one of those decisions per piece. They are the reason a brand books you again.

7Exercise 2.1 — Ten hooks, one product

Take the would replace product from your Module 0B list.

  1. Write the brief. All four required elements. Do not look at the template library yet — write it cold, so you find out which of the four you naturally forget.
  2. Ask for ten hooks.
  3. Read all ten out loud. Mark each one: shoot it / fix it / bin it.
  4. Take your single best one and ask for five more like it, naming what makes it work.
  5. Now open templates/01-testimonial.md and compare your brief to the structured one. The gap is your homework, and it will be the same gap every time until you fix it.

Keep the ten. You'll use the bin pile in Module 4, where showing the takes that failed turns out to be worth more than showing the one that worked.

Your word for it — nothing is tracked automatically.

Source: content/module-2/01-prompting-for-hooks.md