Soku AI
Most AI in advertising is used badly: someone opens a chat window, pastes a screenshot, asks "how do my ads look?", and gets a confident paragraph of generic advice. The output is bad because the input is — no account data, no defined task, no repeatability.
A skill fixes that. It is a recurring ad-ops task written down precisely enough to run the same way every time: what data it reads, what it evaluates, what it outputs, and what it must never do on its own. Same idea as a runbook, but executable.
This page is a library of twenty-four skills worth having, grouped by the job. Each is described so you can implement it in whatever tool you use — a scheduled script, a rule engine, a prompt, or an agentic platform.
Last verified: 2026-08-18.
What makes a skill, rather than a prompt
Four things separate a skill from asking a chatbot a question:
- A defined data input. Not a screenshot — a specific report with named fields and a date range. "Search terms report, last 30 days, with cost, conversions and conversion value" is a skill input. "My account" is not.
- An explicit decision rule. The threshold is written down: spend > $20 AND conversions = 0. If the rule lives in someone's head, the output changes every time it runs.
- A structured output. A list you can act on — ideally one you can paste or apply directly. Prose is not an output.
- A stated boundary. What it must not do without approval. Anything that changes spend belongs behind a human yes.
A skill missing any of these produces advice. With all four, it produces work.
Auditing skills
1. Wasted spend finder. Pull search terms for 30 days; list every term with spend above your threshold and zero conversions. Output: a negative-keyword block, grouped by match type and ad group. The single highest-ROI skill in the library, and the one most often skipped.
2. Full account audit. Structure, budgets, bidding strategies, conversion setup, ad strength, extensions, geographic settings. Output: a graded findings list ordered by estimated cost of the problem. See Google Ads audit checklist.
3. Tracking integrity check. Verify conversion actions are firing, deduplicated, and attributed to the right source. Compare platform-reported conversions against GA4 or the CRM and flag any widening gap. Most tracking failures are found weeks late, by which point the budget is spent.
4. Conversion double-count detector. Specifically checks that browser pixel and server-side API events share an event ID and are being deduplicated. Inflated ROAS from double counting is the most dangerous single error in paid social, because it causes you to scale confidently in the wrong direction.
5. Disapproved and limited ad sweep. List every ad or keyword in a non-serving state with its reason. Rejections inside otherwise-live ad sets distort learning silently.
6. Budget authorisation audit. Every campaign spending money mapped to something someone approved. Finds zombie campaigns from old tests and departed staff.
7. Audience overlap report. Which ad sets are competing for the same people. Self-competition presents as rising CPMs and is invisible unless explicitly measured.
8. Geographic and schedule leak check. Spend outside intended regions or hours. Location-targeting defaults ("presence or interest") send money to the wrong continent more often than anyone expects.
Optimisation skills
9. Budget pacing monitor. Actual versus planned spend by campaign; flag anything over 110% or under 80% with days remaining. Run daily.
10. CPA ceiling enforcement. Identify anything above your maximum acceptable CPA over a stated window with a minimum conversion count. Output: a pause or bid-reduction list. The minimum-volume condition matters — without it this pauses things on noise.
11. Search term harvesting. The inverse of skill 1: converting search terms that are not yet keywords. Output: an add list with suggested match types.
12. Negative keyword hygiene. Conflicts where a negative is blocking a converting keyword. Accumulates quietly over years.
13. Bid strategy fit review. Whether each campaign's bidding strategy suits its conversion volume. Target CPA on a campaign with four conversions a month cannot work; the fix is structural, not a bid change.
14. Learning phase monitor. Ad sets stuck in "Learning limited" and why — usually budget or optimisation event too demanding for the volume. See Facebook Ads checklist.
15. Dayparting analysis. Performance by hour and day of week against a minimum-volume threshold. Output: a schedule adjustment.
16. Placement quality review. Spend by placement with performance. Audience Network and low-quality placements consuming budget is common and easy to exclude once seen.
Creative skills
17. Creative fatigue detector. Frequency rising while CTR falls. Output: which creatives to retire and when. Fatigue is regularly misdiagnosed as an audience problem, and the wrong fix follows. See reach vs impressions vs frequency.
18. Creative performance ranking. Every ad ranked by ROAS, CTR and CPA across platforms, normalised so cross-platform comparison is not nonsense.
19. Test readability check. Given a running creative test, whether it has enough volume per variant to distinguish anything yet. Answers "can we call this?" honestly. See ad creative testing framework.
20. Placement spec compliance. Whether creatives meet size, ratio and safe-zone requirements per placement before launch. Catches the auto-cropped square dropped into Reels.
21. Competitor creative digest. What competitors are running in the public ad libraries, and what changed since last week. Useful as input to concept testing, not as something to copy.
Reporting skills
22. Client or stakeholder report. Spend, CPA, ROAS, top and bottom campaigns, notable changes, next actions — same shape every period. The consistency is the value; a report whose structure changes cannot be compared.
23. Cross-channel roll-up. Blended performance across platforms, with the honest caveat that platform-reported conversions overlap. See MER vs ROAS and cross-channel marketing attribution.
24. Anomaly alert. Statistically unusual movement in spend, CPA or conversion volume against a trailing baseline. Deliberately last, because it is the one people build first and it is nearly useless without the others — an alert with no diagnostic skill behind it just creates noise.
Writing your own
Start from your last five fire drills. Not from a feature list. What went wrong recently that a check would have caught? Those are your first five skills, and they are worth more than any generic library.
Write the decision rule as a number. "Underperforming" is not a rule. "Spend > $20 and conversions = 0 over 30 days" is. If you cannot write the number, the skill is not ready — and the act of writing it usually exposes that the team never agreed on it.
Always include a minimum-volume condition. The most common failure in home-made automation is acting on tiny samples. Every rule that pauses or reduces should require a floor: minimum spend, minimum impressions, or minimum conversions. Without it you will pause your best performer on a slow Tuesday.
Put the boundary in writing. Which skills may act, and which may only report. Our default: anything that changes spend proposes and waits. Read-only skills can run unattended; write skills need a human yes. This is a safety property worth keeping even when the automation is trustworthy, because it preserves the audit trail of who decided what.
Version them. A skill is code even when it is prose. When someone changes a threshold, that should be visible and reversible.
How to run them
Manually, on a checklist — free, and better than not doing it. As scripts, in Google Ads or against the APIs — free, powerful, needs a developer, and needs an owner after the developer leaves. In a rule engine like Optmyzr — auditable and consistent across accounts, but every skill is a configuration project. In an agentic platform — you describe the skill and it reads the accounts and executes on approval, which removes the configuration project but is a newer category with a shorter track record.
The choice matters less than whether the skills are written down at all. A team with twenty-four skills on a checklist and a calendar reminder outperforms a team with an unconfigured rule engine.
Frequently asked questions
What is an AI advertising skill?
A recurring ad-operations task defined precisely enough to run identically every time: a named data input, an explicit numeric decision rule, a structured output, and a stated boundary on what it may do without approval. The definition is what makes it repeatable; the AI is just one way to execute it.
Are these skills free?
The definitions are — every skill here can be run manually from the platform's own reports at no cost. What you pay for is execution: scripts need a developer, rule engines need configuration and a licence, agentic tools need a subscription.
Which skill should I implement first?
The wasted-spend finder. It is the simplest to define, it needs one report, and in most accounts it pays for itself the first time it runs.
Should AI be allowed to change my campaigns automatically?
Our default is no for anything that changes spend. Read-only skills — audits, alerts, reports — can run unattended safely. Write actions should propose and wait, which preserves both a decision point and an audit trail.
How is a skill different from a prompt?
A prompt is a question. A skill specifies its data input, its numeric decision rule, its output format and its authority limit — so it produces the same answer twice given the same account.
