Paid Social Primary Text Ad Creative

Paid Social Primary Text

Shared quality bar: ../references/output-standard.md. All numbers cited here live in ../references/thresholds.md. One angle per run. A batch mixing angles produces copy nobody can read a result from later.

Use this skill when

Feed copy is needed for Meta or LinkedIn on one named angle: primary text, headline and link description, as a set of testable variants. Also use it when existing feed copy buries its point below the truncation line. The angle must already be chosen; this skill writes it, it does not pick between angles. Not for Google search assets tied to keywords, which is search-headline-set-ad-creative. Not for the spoken and on-screen opening of a video, which is video-hook-lines-ad-creative. Not for taking one proven angle and splitting it across cold, warm and retargeting audiences, which is funnel-stage-variants-ad-creative.

Required input

• The angle, written as one sentence naming the argument for acting, per the denominator table in thresholds.md.
• Platform and placement family: Meta feed, Meta reels, LinkedIn feed.
• Audience: who they are and what they already know about the advertiser.
• The offer and the call to action, exactly as the destination states them.
• Every claim the advertiser wants used, each with its source.
• Number of variants required, and the named control they run against.

Analysis workflow

1. Restate the angle in one sentence and get it confirmed. If the supplied angle names a format or an offer rather than an argument, stop and ask.
2. Identify the single idea that must survive truncation. On Meta that is whatever sits before limits.meta_primary_text_visible; on LinkedIn it is whatever sits before limits.linkedin_intro_text_visible. Both cuts move by device, so treat them as where the reader stops rather than as hard limits, and hold LinkedIn headlines to limits.linkedin_headline_target.
3. Write the opening line for each variant first, in isolation. Judge each one with the rest of the copy hidden.
4. Write the body under each opening: the mechanism, then one proof, then the call to action. One proof, not three.
5. Write headlines within limits.meta_headline_visible and link descriptions that add a fact rather than repeat the headline.
6. Apply test.one_change_rule across the variant set. If two variants differ in opening line and offer at once, split them or drop one.
7. Cap the set at test.concurrent_variant_cap and state the named control. Check the ad set volume against test.volume_floor and say plainly whether a readable result is possible at that spend.
8. Apply claims.evidence_rule, claims.superlative_rule and claims.restricted_topics. Mark unsourced factual lines [claim: needs source] and list what was declined.

Decision rules

• The point of the angle appears after limits.meta_primary_text_visible: rewrite, and say which sentence has to move up.
• Variants differing in more than one element: not a test, per test.one_change_rule. Relabel them as separate creatives, not a pair.
• Requested variant count above test.concurrent_variant_cap: write the cap, and explain that the extra variants split budget below test.volume_floor.
• Angle falls inside claims.restricted_topics: write nothing until the advertiser confirms the platform's category requirements. Label approval_needed.
• No proof supplied for the central promise: write the variant without a proof line and say the set is weaker for it, rather than substituting a plausible number.
• Any line this skill produced that carries [claim: needs source] is approval_needed, whatever else its verdict says, and the marker is never stripped to make the set look finished. A marked line is not shippable copy: it waits for the named source, or it is rewritten without the claim. A marked variant is listed apart from the ones that can go live this week, so nobody uploads it by accident.

Output format

No verdict line. One block per variant: primary text, headline, link description, the one changed element versus the control, evidence label, decision label. Then a truncation note stating what each variant says before the visible cut. Then What this could not see, Missing data, Approval gates.

This skill writes copy rather than reading a test or a time series, so it issues no high confidence label. The comparison family's high rests on test.min_runtime and test.volume_floor, which are properties of a test that has not been run on this copy. Confidence here attaches only to the account evidence these variants would later be read on, stated as a forecast about whether a result will be readable, and never as a judgement of the writing.

Practical example

Illustrative and fictional. Advertiser "Foldbrook". Angle: "an in-house team can run paid social without hiring an agency". Meta feed, a variant set requested against a live control at test.concurrent_variant_cap. The ad set delivers about 22 purchases a week, so against test.volume_floor no variant reaches a readable sample inside test.min_runtime. That is stated before the copy, and every downstream verdict on these variants is fixed at low for the volume reason alone.

Variant 1, primary text:

You do not need an agency to run paid social. You need the routine they already run, written down. Foldbrook gives an in-house team the weekly loop: brief on Monday, launch on Tuesday, read the numbers on Friday, decide what to build next. Start free.

Headline: "Run Paid Social In House" (24). Link description: "The weekly loop, not another dashboard." Opening sentence pair is 98 characters, so the whole argument sits above the point where the reader stops.

Variant 2 changes the opening line only, which is the one changed element under test.one_change_rule:

Your agency's real product is a weekly routine. You can run that yourself.

Body, headline and link description are identical to variant 1.

Declined, and reported apart from the launchable set:

• "Agencies charge 4x what this costs." Refused: no agency pricing source was supplied, so claims.evidence_rule and claims.superlative_rule both bite, and it is not rewritten as a vaguer comparison either.
• "Used by 340 in-house teams." Written, but carries [claim: needs source] because the figure arrived verbally, so the variant holding it is approval_needed and does not go up this week.
• A fourth variant built on a named customer's quote. Not written at all: claims.testimonial_permission requires written permission on file and none was supplied, and the permission is the customer's, not the advertiser's.

Two variants can go live this week, one waits on a source, one waits on permission, and the output says which is which rather than handing over four that look equally finished.

Guardrails

• Never write past the point where the reader stops without checking what survives above it.
• Never invent a statistic, price, rating, customer name or comparison.
• Never present variants differing in two elements as a test.
• Never name a winner between variants inside test.material_gap.
• Do not publish, boost or edit anything in the ad account.
• Report every declined variant with the reason, including ones asked for by name.