# Evaluate Ad

> Score a launched paid ad (Meta, Google, TikTok, LinkedIn) from real metrics — network-specific primary metric + diagnosis + fatigue/auction signals — inside an existing eval loop. One cycle per network. Not for loop setup (use run-pipeline), writing new creative (use write-ad), channel-mix retrospectives (use plan-campaign), or campaign-level scoring (use evaluate-campaign).

- **Type:** Skill
- **Install:** `agentstack add skill-hungv47-meta-skills-evaluate-ad`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [hungv47](https://agentstack.voostack.com/s/hungv47)
- **Installs:** 0
- **Category:** [Data & Analytics](https://agentstack.voostack.com/c/data-and-analytics)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [hungv47](https://github.com/hungv47)
- **Source:** https://github.com/hungv47/meta-skills/tree/main/skills/marketing/evaluate-ad
- **Website:** https://skills.forsvn.com/

## Install

```sh
agentstack add skill-hungv47-meta-skills-evaluate-ad
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Ad Eval — Orchestrator

*Evaluation skill. Converts launched paid-ad evidence into a cycle snapshot + ledger row + narrowly-scoped next action inside an existing eval loop. One cycle = one network (+ audience-temp on Meta/TikTok).*

**Core Question:** "Did this ad cycle, on this network, create measurable signal strong enough to keep / discard / watch / block — and what should the next strategy/execution skill know?"

> Why, methodology, history: [`references/playbook.md`](references/playbook.md) [PLAYBOOK]. Capability metadata (route triggers, prerequisites, load map): [`routing.yaml`](routing.yaml).

## Critical Gates

1. **Existing eval loop required.** `program.md` + `context.md` absent → `NEEDS_CONTEXT`, recommend `/run-pipeline`. This skill does not create loops.
2. **Measurement evidence required + the NETWORK's primary metric decides the row.** Need ≥1 metric source, window, current value for the network's primary metric — **Meta** CTR/CPA/ROAS, **Google** Quality Score + impression share, **TikTok** thumbstop/hold rate, **LinkedIn** cost-per-qualified-lead — per the loaded metric surface (`references/ad-intelligence/{google-ads,tiktok-ads,linkedin-ads}-metrics.md`; Meta is inline). Secondary metrics explain diagnosis; don't override unless `program.md` defines a guardrail failure.
3. **One network per cycle (+ one audience-temp on Meta/TikTok).** Networks (and Meta cold/retargeting) evaluated in separate cycles. Mirrors write-ad's one-network-per-artifact. Mixed-network/audience metrics → split before ingest or return `BLOCKED`.
4. **No fabricated analytics.** Unknown stays unknown. Manual notes only when labeled operator-supplied + tied to date/window/source.
5. **Attribution confidence must be explicit.** Every verdict includes sample size (impressions + spend window), baseline comparability, confounders (creative change mid-flight, audience shift, iOS attribution gap), confidence: `high | medium | low | blocked`.
6. **Evaluation does not generate creative.** Recommend next changes; route creative authorship to `write-ad` (revised brief), channel-mix to `plan-campaign`, LP-bottleneck to `brief-landing-page`.

## Responsibility Split

`/run-pipeline` owns loop setup + `program.md` / `context.md` / `results.tsv` schema + durable learnings. **This skill** owns post-launch Meta-ad evidence snapshots for a loop cycle, scoped to a single audience-temp. `/write-ad` owns next-cycle creative.

## Inputs

**Required:** loop slug/path · network (`meta | google-ads | tiktok-ads | linkedin-ads`) · audience-temp tag (Meta/TikTok: `cold-traffic` OR `retargeting`) · source ad-copy artifact (`docs/forsvn/artifacts/marketing/write-ad/[audience-temp]-[date]-[slug].md`) · measurement window · the network's primary metric value + source · spend window.

**Recommended:** baseline/prior cycle row · frequency at window close (Creative-Fatigue scoring) · conversion count + CPA · audience size + reach · guardrails from `program.md` · qualitative evidence (comments, sentiment, click-quality).

## Outputs

`.forsvn/loops/[slug]/evals/YYYY-MM-DD-cycle-N.md` + append one row to `results.tsv` via `bun scripts/append-loop-result.ts` (8-col helper) + update `learnings.md` ONLY for high-confidence keep/discard reusable lessons (critic-gated) + run `bun scripts/manifest-sync.ts`.

## Agent Manifest + Dispatch

4 sub-agents: Layer 1 parallel (Metric Ingest + Diagnosis) → Layer 2 (Recommendation) → Layer 3 (Critic). Critic FAIL → revise once; still FAIL → no ledger row + `BLOCKED`. Full agent table + per-layer dispatch + 7-dim rubric: [`references/agent-manifest.md`](references/agent-manifest.md). Domain rubric: [`references/rubric.md`](references/rubric.md). Shared frame: `_shared/evaluation-loop-rubric.md`.

## Pre-Dispatch

Canonical: `_shared/pre-dispatch-protocol.md` + `_shared/eval-loop-spec.md`. **Hard-blocks (BEFORE Cold Start):** missing `program.md`/`context.md` → `NEEDS_CONTEXT` + `/run-pipeline`; no measurement evidence → `BLOCKED`; mixed-audience metrics with no clean split → `BLOCKED`; custom 10+ col `results.tsv` → warn + hand-edit. **Cold Start:** 6 bundled questions (loop · audience-temp · source ad-copy path · window · primary metric value/baseline · spend window). Full read-order + warm/cold templates + `--fast` behavior: [`references/procedures/pre-dispatch.md`](references/procedures/pre-dispatch.md) [PROCEDURE].
Session execution profile (single-vs-multi): inherit per `references/_shared/execution-policy.md`.

## Artifact Contract

- **Path:** `.forsvn/loops/[slug]/evals/YYYY-MM-DD-cycle-N.md`. **Lifecycle:** `evaluation`.
- **Frontmatter (10 fields):** `skill` / `version` / `date` / `status` / `summary` / `purpose` / `lifecycle` / `use_when` / `do_not_use_when` / `upstream` / `downstream` + provenance (`input_artifacts` = source ad-copy + `brand/BRAND.md` + `research/icp-research.md`; `output_eval: null` — eval cycles are typically terminal).
- **Body sections (8):** Title · Verdict · Evidence (6-col table) · What Changed This Cycle · Diagnosis (Likely Drivers + Confounders + Creative-Fatigue Signals) · Next Cycle Recommendation · Results Row (8-col TSV) · Learning Promotion. Verdict block must name the audience-temp (Gate 3); Evidence table scopes metrics to that audience-temp.
- **Results Row (8 cols):** `cycle  date  artifact  primary_metric  value  baseline  status  description`. `status` ∈ `keep|discard|watch|blocked` (Critic Hard Fail otherwise); description includes the audience-temp tag.
- **Cross-stack contract:** consumed by future ad-eval cycles (trend analysis) + `write-ad --rev=N+1` (hypothesis seeding) + human reviewers. Ad-eval does NOT directly consume write-ad output — sibling coordination is at the eval-loop boundary. Schema changes require atomic update across `_shared/eval-loop-spec.md` + downstream callers.

Full byte-identical template + Evidence/Results/Learning formats + helper invocation: [`references/format-conventions.md`](references/format-conventions.md) [PROCEDURE].

## Results Row Helper

```bash
bun scripts/append-loop-result.ts "" \
  --artifact evals/YYYY-MM-DD-cycle-N.md \
  --metric "" --value "" --baseline "" \
  --status "" --description ""
```

Do not append on Critic FAIL — return `BLOCKED` instead.

## Critic Override Protocol

Operator ships despite critic FAIL — **log BEFORE writing artifact or ledger row:** `bun scripts/log-critic-override.ts --skill evaluate-ad …`. Three overrides → rubric-revision escalation. An override never promotes a contested cycle to `keep`; a no-override FAIL still returns `BLOCKED`. Full protocol: [`references/_shared/critic-override-protocol.md`](references/_shared/critic-override-protocol.md) [PROCEDURE].

## Anti-Patterns

[`references/anti-patterns.md`](references/anti-patterns.md) [ANTI-PATTERN] — 7 ad-eval rows + 4 cross-cutting marketing-stack. Most common: mixed-audience contamination (Gate 3 + Critic "Audience-Temp Fidelity"), confidence inflation on low-spend (Gate 5 + Critic "Attribution Honesty"), scope drift to write-ad (Gate 6 + Critic "Decision Discipline"), missing source ad-copy artifact (Critic Hard Fail).

## Worked Example

Ad-eval cold-traffic cycle (creative-fatigue keep→watch, critic PASS): [`references/examples/ad-eval-cycle-walkthrough.md`](references/examples/ad-eval-cycle-walkthrough.md).

## Durable Rules (protected)

## Completion Status

- **DONE** — eval artifact written, ledger row appended, critic PASS.
- **DONE_WITH_CONCERNS** — artifact + row written, but confidence low/medium or confounders material.
- **NEEDS_CONTEXT** — missing loop, source ad-copy artifact, audience-temp tag, or required metric evidence.
- **BLOCKED** — contradictory data, mixed-audience ingest, filesystem failure, or critic failed after revision.

## References

- `references/{playbook, agent-manifest, rubric, format-conventions, anti-patterns}.md` + `procedures/{pre-dispatch, dispatch-mechanics}.md`
- Per-network metric surfaces (loaded by `network`): `references/ad-intelligence/google-ads-metrics.md` · `references/ad-intelligence/tiktok-ads-metrics.md` · `references/ad-intelligence/linkedin-ads-metrics.md` (Meta metrics stay inline in the agents).
- `_shared/{eval-loop-spec, evaluation-loop-rubric, pre-dispatch-protocol, critic-override-protocol, quality-dashboard-spec}.md`
- **Siblings:** `write-ad` (downstream — this skill routes recommendations TO it but does not produce briefs), `run-pipeline` (loop scaffolding + ledger schema + durable learnings), `evaluate-campaign` (multi-channel aggregate sibling)

## Source & license

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

- **Author:** [hungv47](https://github.com/hungv47)
- **Source:** [hungv47/meta-skills](https://github.com/hungv47/meta-skills)
- **License:** MIT
- **Homepage:** https://skills.forsvn.com/

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

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-hungv47-meta-skills-evaluate-ad
- Seller: https://agentstack.voostack.com/s/hungv47
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.