Write Ad

Write Ad — Orchestrator

Ready-to-publish paid-ad copy across Meta, Google, TikTok, and LinkedIn — one network per artifact, network-aware caps + framing. Multi-agent strategy → draft → format → voice → critic → humanmaxxing pipeline. Capability metadata lives in [routing.yaml](routing.yaml). Agent table + 2 routes + 7-dim rubric + post-humanmaxxing regression rule: [references/agent-manifest.md](references/agent-manifest.md). Methodology: [references/playbook.md](references/playbook.md).

Core question: Would this ad still make sense if the platform stripped every claim that isn't substantiated by a named entity or measured number?

Critical Gates — load first

1. Network + audience-temp are non-negotiable. network (meta | google-ads | tiktok-ads | linkedin-ads) selects the format-checker spec + the strategist's framing axis; ambiguous network BLOCKs — ask one question, never assume Meta. Missing audience-temp BLOCKs. Offer + creative-format + proof also hard-block.
2. One network + one audience-temp per artifact. Cross-network or cross-temp campaigns run the skill once per pair; do NOT stack two networks' char-caps/policy in one artifact.
3. Hero + 2 distinct variants per artifact is the default. Strategist enforces 3 distinct angle_archetype values + 3 distinct anchor_proof entries. Variants isolate ONE variable each (Variable Subtraction). The count serves Meta A/B structure; an operator may override it — log the override + reason in the rationale. Mode + tier in [references/_shared/options-selection.md](references/_shared/options-selection.md) (DELIVERY, tier B).
4. Format-checker is a HARD gate, not a critic dim. Bounces on the resolved network's char-cap violation / banned policy phrase / unsubstantiated measured claim. PASSED / REVISIONREQUIRED (does NOT consume critic cycle) / FORMATFAIL (escalate to user).
5. Humanmaxxing runs ONCE per variant with protected_tokens including URL. Post-humanmaxxing Specificity regression check per variant — drops ≥2 OR named entity/number/URL absent → flag the delta + the removed specific for operator review (critic-approved draft preserved alongside; operator picks).

Quality Gate — 7 dimensions

Full rubric + Pass/Fail + per-dim scoring bands: [references/agent-manifest.md](references/agent-manifest.md) § 7-Dim Critic Rubric. Domain rubric: [references/rubric.md](references/rubric.md).

• Gate: Total ≥49/70 AND every dim ≥6. Total 49-55 with all dims ≥6 = PASS as DONE_WITH_CONCERNS. Any dim 30% from peak) OR offer change OR LP change.

Assisted/Direct: a verified engine submits/renders, you approve at the gate. See [execution-fork.md](references/_shared/execution-fork.md); record execution_mode.

Worked Example

End-to-end Route A (MealKit cold-traffic 14-day-trial app, 3 distinct variants: outcome-first / scale+social-proof / authority+mechanism, format-checker REVISIONREQUIRED on char-cap + health-claim disclaimer, composer revised, critic PASS aggregate 168/210, terminal humanmaxxing per variant with protected_tokens incl. URL, post-humanmaxxing Specificity regression passes per variant) + cycle-2 FAIL variant + Format-Checker REVISIONREQUIRED path + Route B snippet: [references/examples/write-ad-walkthrough.md](references/examples/write-ad-walkthrough.md).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hungv47
• Source: hungv47/meta-skills
• License: MIT
• Homepage: https://skills.forsvn.com/

Install and usage instructions live in the source repository linked above.