---
name: "ad_angle_engine"
description: "Turns one product into a full ad campaign: expands 7 angle families (pain, desire, proof, contrarian, mechanism, identity, urgency) into written ad variants, screens every line against platform policy, and ships a ranked test plan with kill and scale criteria. Triggers when someone needs ad copy, ad concepts, or a creative testing plan for Meta, Google, LinkedIn, or TikTok."
metadata: { "includeInPrompt": true }
---

# Ad Angle Engine

## Purpose

One strong offer, seven different reasons to care. This skill takes a product (or an offer) and produces a campaign's worth of ad copy built on a real angle system: each angle attacks a different buyer motivation, gets a written hook line, survives policy screening, and enters a ranked test plan with clear decision rules. The point is not variety for its own sake. It is a calibrated set of testable bets, so the account learns which motivation actually moves buyers.

## Workflow

1. **Extract the grounded offer.** Before writing a single word, pin down: product, core promise, target buyer, platform, ad format, landing page. If the core promise is fuzzy, stop and say so. Twelve angles on a fuzzy offer produce twelve fuzzy ads. Nothing proceeds on assumptions: every number, testimonial, and result quoted must come from the user or a source they provide. Unverifiable claims get cut, not softened.

2. **Expand the angle families.** Build at least one angle per family in `references/angle-families.md` (pain, desire, proof, contrarian, mechanism, identity, urgency). Each angle gets:
   - A one-line angle statement: the specific reason a buyer would care.
   - A written hook line: the actual first line a person would see, in the platform's format (ad headline, primary text opener, search ad headline). An angle without a hook line is not testable and does not ship.
   - A distinctness check: if you deleted the other six, could this one stand as its own ad? Synonym swaps fail the check. Rewrite until it is genuinely different.
   - Skip a family only with a written reason (e.g. urgency does not fit an evergreen subscription). Note it in the output.

3. **Flesh out full variants.** For each surviving angle, write the full ad: headline(s), body/primary text, CTA, visual concept. Keep to the platform's character and format constraints; verify current official limits before shipping exact compliance (the reference values in `references/policy-checklist.md` are starting points, not gospel).

4. **Screen every line for policy.** Run each variant through `references/policy-checklist.md`. The big ones: Meta's personal-attributes rule (no direct or indirect assertion about age, health, finances, religion, identity, or any of the protected categories), no appearance-inferiority copy, no unqualified time-boxed outcome promises, evidence qualifiers on every result claim. Rewrite or drop anything that fails. A rejected ad costs more than a cut angle.

5. **Validate with the checker.** Put the variants into a CSV bank (angle, platform, element, copy, claim_source) and run `python3 bin/bank_check.py bank.csv`. Fix every FAIL before continuing. This is mechanical hygiene, not a substitute for the policy read in step 4.

6. **Build the ranked test plan.** Rank the variants and lay out the test per `references/test-plan-rubric.md`: which variants run first, the budget per variant, the minimum data before any read, the kill and scale criteria, and the review cadence. One meaningful variable per test cell: angle changes, offer and landing page stay fixed. Ship the plan as part of the output, not as an afterthought.

7. **Iterate from data (when performance data exists).** Compare against the decision metrics in the rubric. Name winners by the agreed metric, keep losers paused (not deleted; a loser for one audience can win for another), and brief the next wave from what won: variants of winners, not new random ideas.

## Output Contract

Deliverables, in this order:

1. **Angle bank** (markdown or CSV): angle family, angle statement, hook line, full copy with character counts, policy status per variant, claim sources.
2. **Test plan**: ranked variant order, budget per variant, minimum-data thresholds, kill/scale criteria, review dates.
3. **Iteration report** (only when performance data exists): evidence, hypotheses, pause/scale/test recommendations.

Table format for the bank:

| Angle | Platform | Element | Copy | Chars | Policy | Claim source |
|---|---|---|---|---|---|---|

Every factual claim in every row must have a claim source, or the claim is cut.

## Operating Rules

1. No invented numbers, dates, quotes, or events, ever. A stat without a source does not exist.
2. Result promises always carry an evidence qualifier: "on average", "based on", "up to". Never promise what the landing page cannot fulfill.
3. No empty superlatives, no all-caps, no fake scarcity, no rhetorical questions aimed at personal attributes ("Still struggling with your weight?" is a policy failure, not a hook).
4. Every search-ad headline must stand alone and combine safely with any other headline.
5. One variable per test cell. Never test a new angle and a new offer in the same cell.
6. Never call a winner below the rubric's minimum data. Variance at small samples looks like signal and is not.
7. Keep SKILL.md tight; family recipes live in `references/angle-families.md`, policy in `references/policy-checklist.md`, test discipline in `references/test-plan-rubric.md`.