Skip to content
Primores AI wiki

Fully Prescriptive Production Briefs for AI-Assisted Content Teams

Fully Prescriptive Production Briefs for AI-Assisted Content Teams

TL;DR: When production is AI-assisted and cheap, creative decisions are the bottleneck and the risk — so the brief must carry every decision: final copy, char-counted overlays, frame-level slot tables, exact CTAs. Executors flag problems; they never invent. Quality bar: a stranger with the assets and zero context could produce all variants identically twice. The two structural moves: hooks live in slot 1 only (one base edit, three swapped openings, each a different mechanism), and every visual row carries an asset-source tag (AI-gen / shoot / screen-rec / stock / brand-asset / UGC-cleared) that routes it into the right production pipeline.

The principle: prescriptiveness follows from cheap production

Traditional briefs are suggestive (“convey energy, maybe show the product in use”) because human craft fills the gaps and production is expensive enough that few variants get made. AI-assisted production inverts both conditions: variants are nearly free, and the executors — human editors, image models, video models — fill gaps with inventions rather than craft. A vague brief at AI speed produces 100 plausible, mutually inconsistent, untestable creatives.

So the brief becomes fully prescriptive: every visual described well enough to hand verbatim to a camera operator or an image model, every word final, overlay copy char-counted, CTA labels exact. “Suggestion” language (“could,” “maybe,” “consider”) in a final copy field is a QA failure. The posture is executors flag, never invent — a blocked dependency gets marked BLOCKED with a buildable alternative offered, never silently substituted.

The test that makes the bar concrete: the package is done when a stranger with the assets and zero context could produce all variants identically twice.

This is glossary/automation-eats-execution applied to the brief itself: AI compresses execution, which moves all the leverage (and all the risk) into the decision layer — and the brief is where the decision layer is serialized.

Hook = slot 1 only

The hook — the first 0–2 seconds of video, the overlay layer of a static — is the main performance lever. The data behind that is old but official: Facebook IQ/Nielsen found 47% of a video campaign’s value lands in the first 3 seconds (2016); TikTok’s creative guidance puts ~90% of ad-recall impact inside the first 6 seconds. Practitioners track it as “hook rate” (% of impressions reaching 3 seconds).

The structural consequence: one base edit per video, three swapped openings. Nothing downstream of slot 1 varies. The industry calls the general practice modular creative; the discipline here is stricter than most modular setups in two ways:

  1. Hooks must differ in mechanism, not wording. Three synonyms of the same pain-callout is a QA failure — it ruins the test read. The working mechanism taxonomy: pain-callout · stat-shock · curiosity-gap · identity-call · objection-flip · social-jolt · honest time-urgency · contrast/before-state · question-trap · pattern-interrupt. Mechanisms also get spread across creatives so a package doesn’t over-index one opening type.
  2. Hooks are seeded from harvested customer language (the unit’s hook seeds from marketing/evidence-graded-audience-research) — adapted in emotional shape and vocabulary, never quoted as testimonials (see guardrails).

The payoff is cheap variants with clean test reads: 6 base creatives × 3 hooks = 18 named ad variants per unit, where a winning variant tells you which mechanism won.

The card formats

Each unit ships as 3 static cards + 3 video cards. The fields that do the load-bearing work:

  • Video slot table — frame-level rows: time range, visual description, final VO line, final on-screen text, asset source. Slot 1 is the hook slot; slots 2–N form the base edit (verified to play correctly after all three hooks).
  • Static layout — fixed composition (“what is shown” stable across hooks), hook zone top third, platform safe zones, hooks as headline (≤8 words) + subline (≤12), char-counted.
  • Destination block — exact URL, which offer is sold, and a message-match note (the landing page must answer the hook). Gaps between ad message and live landing pages are emitted as a deliverable for the LP team, never left as silent mismatches.
  • Proof-requirement coverage map — which card/slot delivers each of the unit’s proof requirements; every requirement must be delivered by at least one card.
  • Reference link — mandatory per card; no reference, no card.

The asset-source column

Every visual row carries one tag — and the tag is the handoff selector between brief and production pipeline:

TagRoutes to
AI-genImage/video generation — a generation-ready prompt is mandatory in the card (one-liners fail QA)
shootThe aggregated shotlist (below)
screen-recApp/report/site capture, with the exact screen state specified
stockLicensed stock, with search terms + fallback
brand-assetClient inventory references
UGC-clearedReal user content with named, cleared rights

All shoot rows flow into one shared, deduped shotlist written for cross-unit reuse: “woman 35–45, neutral bathroom, examines shelf of products” serves three units; “holds the report, page 3” serves one. Prefer the former phrasing wherever content allows — the next wave reuses footage by design. In the reference batch, 60 videos’ worth of shoot rows collapsed into five setup groups.

What a real batch taught: hook collisions

The reference batch — 20 units → 120 base creatives → 360 named variants, written by parallel agents off one central briefing file with centrally allocated concepts — surfaced a non-obvious failure mode in its mandatory similarity audit:

Hook collisions happen at the copy level even when concepts are centrally allocated. One sourced statistic opened hooks in four different units’ creatives; two sibling units ran a near-identical question-trap opening. Central allocation prevented concept duplication, but four different writers independently reached for the same strongest stat as an opener — and if four test cells open with the same stat, the test read is contaminated: you can’t tell which angle won.

The rules that came out of it:

  • Signature hooks are exclusive to one cell. The stat-as-hook belongs to the unit whose angle it is; other units may use the stat in body slots (sourced), never as the opener.
  • Similarity audit-then-fix is a mandatory batch stage, not an optional polish: review all hooks and all opening visuals across the batch before anything ships. (Adjacent practitioner corroboration: Meta’s Andromeda-era guidance pushes distinct concepts over near-duplicate variants, since the delivery system recognizes and suppresses redundancy — directionally aligned, though the platform-side numbers circulating for it are unverified.)

Guardrails worth stealing

  • Harvested quotes are never testimonials. Hook seeds are real people’s public words — adapt the language and emotional shape; never present them as customer testimonials, imply product use, or fabricate reviews. This is FTC territory: the 2023 Endorsement Guides make republished third-party praise an endorsement the advertiser is liable for, and the 2024 Consumer Reviews rule bans fake or misrepresented testimonials with civil penalties (~$51K/violation). Real testimonials require cleared rights, marked in the asset-source column.
  • Organic trending sounds are not licensed for paid ads. Cleared paths only: TikTok’s Commercial Music Library or Meta’s Sound Collection, or original/VO-only audio. Specifying a trending sound in a paid brief is a QA failure.
  • Ad names encode unit IDs (TA2J1A1-V2-B = unit, video 2, hook B) — non-negotiable, because the research layer’s feedback merge keys on the names. A media buyer’s results sheet maps back to units automatically, with hook-level vs angle-level failure attribution intact.
  • Compliance flags travel with the card. If the research run couldn’t verify compliance, every card carries compliance: unverified — the buyer decides exposure consciously instead of inheriting an invisible risk.
  • Brand coherence is an input, not an accident: a run-level brand spec (colors, fonts, product-render references, voice adjectives) is distilled once and referenced by every card, so parallel executors converge on one brand without guessing.

Honest limits

  • One full batch behind the method (anonymized: a DTC genomics-skincare engagement). The hook-collision finding is real but N=1 at batch scale; the guardrails and formats are anchored in public rules and platform docs.
  • The hook-window numbers are the field’s weakest-but-best data: the 47%-in-3-seconds figure is from 2016, and the widely-circulated “71% decide in 3 seconds” TikTok stat has no traceable primary source (we don’t use it).
  • Fully prescriptive briefs cost real authoring time. The trade is deliberate: spend it once at the decision layer, save it at the (cheap, parallel) execution layer. For a one-off creative, a conventional brief is fine.

Key Takeaways

  • When production is cheap, the brief is the product: every visual described, every word final, char-counted, executors flag but never invent.
  • Hook = slot 1 only: one base edit, three openings, three different mechanisms — synonym hooks ruin the test read.
  • The asset-source column is the routing layer between creative decisions and production pipelines; shoot rows aggregate into a deduped, reuse-first shotlist.
  • Hook collisions happen at copy level even with central concept allocation — make signature hooks cell-exclusive and run a similarity audit before launch.
  • Harvested quotes ≠ testimonials (FTC), trending sounds ≠ paid-licensed, and ad names must encode unit IDs so performance feedback maps back automatically.

Sources