Skip to content
Primores AI wiki

Target Audience Research — Research-to-Units Compiler for Paid Social

Target Audience Research

TL;DR: A staged research compiler for paid-social creative strategy: brand inputs of any depth go in, an evidence-graded unit table [TA × JTBD × angle + content type] comes out — every claim labeled provided | researched | hypothesis, every audience built from verbatim customer language, priority scored with transparent sub-scores, IDs frozen at sign-off. Three human gates (research scope, TA confirmation, final sign-off). Depth-adaptive: rich client input means thin research; thin input means heavy research. The methodology it implements is documented at marketing/evidence-graded-audience-research.

What It Does

The skill replaces the persona deck with a compiler. You hand it whatever exists — an offer doc, pricing, asset lists, links, or almost nothing — and it produces the machine-readable strategy layer that creative production can actually consume: ranked, evidence-graded units with stable IDs.

The pipeline, end to end:

StageWhat happensGate
0 — IntakeClassify every brief field ask-only / file-preferred / researchable; ONE question batch with skip-consequences—
1 — Research planScope research modules by depthGate 1: scope approval
2 — ResearchParallel agents: competitive ad-angle sweep + community/review language harvest, all access-dated—
3 — Target audiencesTAs assembled from verbatim language, each with an awareness levelGate 2: TA confirmation
4 — JTBD tournament3–5 job candidates per TA → keep 1–3, rejects listed with the test that killed them—
5 — Angle tournament3–6 candidates per [TA×JTBD] slot, five named checks (mechanism, awareness, sophistication, persuasion lever, saturation)—
6 — Unit tableStable IDs, priority sub-scores (evidence/whitespace/feasibility/economics), wave-1 flagsGate 3: sign-off → IDs freeze

After sign-off, the table is append-only: field results merge back (the skill detects ad-platform exports and wave-results framing), get attributed at hook/angle/TA level, and re-rank the queue — new learnings become new IDs, never renumbered ones.

What Makes It Different

  • Refuses to fabricate. Ask-only fields (margins, compliance, capacity) are never researched — a skip becomes a labeled hypothesis with its downstream consequence recorded. Model training knowledge may suggest search queries; nothing survives into the output unless fetched this run.
  • Verbatim capture as a hard rule. Every pain/desire/objection quote is stored in original language + translation + source link. Paraphrase destroys the value — quotes become hook seeds downstream.
  • Tournaments with visible rejects. Rejected job and angle candidates are listed per slot with the specific check that killed them. The anti-whitespace-bias rule forces one saturated-but-winnable candidate per slot.
  • Frameworks are load-bearing, not decorative. Schwartz awareness + sophistication decide angle forms; each angle names its Cialdini lever; marketing/discovery-before-scale shapes the priority queue into a testing queue.

Real Run: The Engagement Dogfood

First full run (a DTC genomics-skincare brand, US market, June 2026): intake → one question batch → 5 TAs confirmed → 8 JTBD → 11 angles → 11 units signed off, wave-1 = 6.

What the run surfaced that the client’s own inputs didn’t:

  • A competitor listed as live in the intake file was dead (domain redirecting to a retailer) — fresh-fetch discipline caught what the file rot hid.
  • The US paid lane was emptier than the input research implied — with two dead premium players as the caveat (whitespace ≠ easy). Published as competitor-analysis/dna-beauty-paid-social-whitespace.
  • A pricing collision: the client’s subscription price sat 4–6× above the churn-trigger line documented in analog-brand reviews → became a binding creative rule (anchor vs treatments, not products).

Honest limits from the run: no ad-library access in-session meant the competitive sweep ran indirect (lander inference + teardowns) — whitespace conclusions carry a medium-confidence cap; compliance fields were skipped by the client, so every downstream brief carries compliance: unverified.

When to Use

  • “Target audience research / segmentation / avatars for this brand”
  • “Who should we target and with what message?”
  • “JTBD / core promise / ad angles research”
  • Kickoff research for a new paid-social client — at any input depth

Chains With

Upstream of tools/scenario-compiler (consumes the signed-off unit table); uses tools/reddit-thread-analyzer-style community mining and tools/pdf-streamer for long source docs; leverages the wiki’s framework pages as priors.

Key Takeaways

  • Inputs are classified by who can know them; research never substitutes for asking, and skips become labeled hypotheses.
  • The deliverable is a unit table with stable IDs and transparent sub-scores — not an avatar deck.
  • Tournaments generate wide and cull visibly; the saturated-but-winnable rule prevents whitespace bias.
  • Three gates hold the irreversible decisions; everything between them runs at machine speed.
  • Field results merge back at hook/angle/TA level — research compounds across waves.

Sources

  • Primores internal skill (14-target-audience-research: SKILL.md, pipeline stages, brief schema)
  • First engagement run, June 2026 (internal; client anonymized)