Paid Ads — Build, Diagnose, Decide — Output First

A working playbook for performance marketers across Google, Meta, LinkedIn, TikTok, and X. Routes any paid ads request into the right task flow, delivers a structured output immediately — before asking clarifying questions — and covers strategy, diagnostics, budget decisions, reporting, and platform evaluation. No intake questionnaire. No analysis-paralysis. The deliverable comes first; the refinement questions come after.

When to Use This

1. Your ROAS dropped, CPA spiked, or CVR fell and you need to know why
Performance degraded over the last 7–14 days and you need a ranked list of what's most likely wrong — not a generic checklist. Use this for diagnostic work: the skill maps your symptoms to a decision tree (auction-side vs creative decay vs site/tracking issue), produces a ranked hypothesis list with confidence levels and verification methods, and separates ""fix the tracking"" from ""fix the campaign"" from ""fix the landing page.""

2. You're building a new campaign, entering a new market, or setting up a new account from scratch You need a platform recommendation, an account structure, a naming convention, and a launch-gating checklist before you spend a dollar. Use this when you're starting from zero or rebuilding: it infers your goal from whatever context you provide, selects the right 1-2 platforms, designs the campaign hierarchy, allocates budget across validation and scale phases, and runs a pre-launch check.

3. You need to make a scaling, kill, or budget-shift decision
ROAS is at 2.4 — should you scale? One campaign is underperforming — kill it or wait? You want to shift budget from Google to Meta. Use this when the decision is on the table and you need a direct Yes/No/Wait with a rationale, a flag if data is too thin to be confident, a structural risk warning (e.g., this change resets the learning phase), and a kill rule if one doesn't exist.

4. You need a report to send your boss or client
Monday morning, weekly review, end-of-month recap, board update. Use this to produce a structured retrospective: three-sentence headline, metric table with source-of-truth vs platform-reported split, root causes ranked by confidence, actions with owner and deadline, and a mandatory re-evaluation date. Designed so the reader gets exactly four things: what happened, why, what you're doing about it, and when to check back.

5. You're evaluating whether to enter a new platform or make a strategic bet
Should we try TikTok? Is our $180 CPL actually good or bad? Should we increase total ad spend? Use this for strategic evaluation: audience fit and creative format demands on the target platform, minimum viable test budget for statistical signal, and how to contextualize benchmarks against your own funnel economics (not industry averages in isolation).

How It Works

Every request is routed to one of five task flows before any output is produced. The routing determines the framework applied and the shape of the output.

The five task flows

Build — New account, new market, or new campaign from scratch

1. Infer goal and break-even from available context (state assumptions explicitly if missing)
2. Select 1-2 platforms from a matrix of business model × intent type × audience reachability × creative capacity
3. Design account structure: Campaign [objective/stage/geo] → Ad set [targeting slice] → Ads [3-6 creative variants]
4. Allocate budget: 70/30 split between baseline and structured testing for the first 2-4 weeks; consolidate once a winner has 30-50 conversions
5. Run pre-launch gate: verify tracking with a real test conversion before spending

Diagnose — ROAS dropped, CPA spiked, CVR fell

1. State the interpreted version of the metric, window, and baseline in one line
2. Run the diagnostic decision tree: top-level split (cost-side vs creative/audience vs site/tracking) → ranked branch hypotheses
3. Flag any data quality issues as ⚠ caveat — not a blocker on the diagnosis
4. Produce: primary hypothesis + confidence level + verification method; secondary hypothesis; recommended next actions ordered by cost-to-test

Decide — Scale, kill, swap creative, shift budget

1. State the decision immediately: Yes / No / Wait + one-line rationale
2. Flag data sufficiency (below 30 conversions on a CPA/ROAS decision = noted, not blocking)
3. Flag structural risk if the change resets learning (budget moves >25%, event changes, attribution window changes)
4. Set a kill rule if one doesn't exist
5. Output: decision + rationale + expected result + date to re-check

Report — Recap, weekly/monthly review, client update

1. Three-sentence headline (hard limit)
2. Metric table with both platform-reported and source-of-truth revenue when they diverge >15%
3. Root causes ranked by confidence with supporting and contradicting evidence listed
4. Actions table: priority / action / owner / start / due / expected effect
5. Re-evaluation section: leading indicators (visible 3-5 days), lagging indicators (7-14 days), next mandatory checkpoint, escalation rule

Advise — New platform evaluation, benchmark contextualization, strategic bet

• Platform evaluation: audience reachability and demographic fit, creative format demands, signal quality (can conversions be tied back?), minimum viable test budget
• Benchmark evaluation: CPL/ROAS/CPA in context of downstream funnel economics (lead → SQL → close rate), not industry averages in isolation

The output-first rule

Every response produces a concrete deliverable first, using whatever information is available. Missing information is handled by making a reasonable assumption and stating it inline. Clarifying questions appear at the end — limited to the 2-3 most load-bearing for the specific task. The intake questionnaire never appears before the output.

Platform-Specific Defaults

Default starting points per platform — adjusted based on business model and creative capacity:

Most accounts perform best on 2 platforms. Fragmenting budget across 5 platforms is one of the most common structural failure modes.

What Makes This Different

vs asking a general AI for paid ads help
General AI starts with clarifying questions. You answer 8-10 of them, then receive generic advice that doesn't commit to a specific recommendation. This skill produces the deliverable first — structure, diagnosis, decision, or report — with stated assumptions, and follows up with the 2-3 questions that would most sharpen the next iteration.

vs using platform-native tools (Google Recommendations, Meta Advantage)
Platform tools optimize for platform metrics. They treat platform-reported ROAS as ground truth. This skill separates platform-reported numbers from source-of-truth data (Shopify/CRM), flags divergences >15% as a tracking caveat, and distinguishes between a campaign problem, a tracking problem, and a landing page problem — which platform tools are structurally incentivized not to do.

vs a paid media agency playbook
Agencies produce the same report template regardless of the situation. This skill routes to the specific task type that matches the actual question, which means a scaling decision gets a Yes/No/Wait with a kill rule — not a 12-page deck. A diagnostic gets a ranked hypothesis tree — not a generic performance review.

The structural outputs that generic advice doesn't produce

• Retrospective reports with mandatory owner, deadline, and re-evaluation date — not action items that dissolve after the meeting
• Diagnostic hypotheses ranked by confidence with stated verification methods — not a checklist of things that ""could"" be wrong
• Decision outputs with explicit data sufficiency flags and learning-phase reset warnings — not opinions dressed as facts
• Pre-launch gating checks tied to actual tracking verification — not a reminder to "set up pixels"