Why Claude gets its own list
Claude connects to Google Ads through a native custom connector — five clicks, no passwords — and from that moment its answers come from your live spend, search terms and conversions instead of from general knowledge. (If MCP is new to you, the MCP server explainer covers what the connection actually is.) These fifteen templates assume that setup, and three of them use things only Claude has: projects that hold your account context permanently, a context window long enough to swallow a whole 90-day export in one paste, and artifacts that turn a report into a document you can send.
If you work in ChatGPT instead, the 37-prompt ChatGPT library is the same job for that assistant.
Two tags mark every prompt. Needs account access means it asks about your data — unconnected, Claude produces a polite guess. Works in any chat means it is about writing, structure or reasoning. Each template ends with what a good answer looks like, because that line is how you tell a real answer from a fluent one.
Set up once
1. Project instructions that every prompt inherits
Add this to the project instructions, edited to fit: You manage Google Ads for [business, one sentence]. Market: [countries, language]. A conversion is [action] and is worth roughly [value]. Treat any search term with more than [20] in spend and zero conversions as waste worth flagging. Standing rules for every conversation in this project: - always state the date range you used - always show the previous period next to the current one - never estimate search volume or CPC; say where I can get them - propose account changes, never make them without my explicit yes
Works in any chat.
This is the one genuinely Claude-shaped move on this page. Paste it into a project's instructions once and every conversation opened inside that project starts already knowing your business, your thresholds and your safety rule — which is why the templates below can stay short. A good sign it is working: Claude names your conversion action unprompted.
2. The connection check
List the Google Ads accounts you can currently see, and for the main one tell me the most recent date you have complete data for. Answer with names and dates only — no analysis yet.
Needs account access.
Run this before trusting anything else. A good answer is boringly specific: account names, customer ids, a date. If Claude answers in generalities or asks you to paste data, the connector is not actually attached to the conversation — the setup guide's troubleshooting section covers why.
Find the leak
3. The waste ranking
Using the last [90] days, rank where this account wastes money. Count as waste: search terms and keywords with spend and zero conversions, ad groups with cost per conversion above [2]x the account average, and spend in geographies or on devices converting far below average. State your definition before the numbers. Show a ranked table with spend attached, total it, and say which line you are least sure of. Recommend nothing yet.
Needs account access.
The definition-first rule exists so you can disagree with what was counted. If the total arrives without a least-confident line, ask for one — a waste figure with no stated uncertainty is a headline, not a finding.
4. What changed, and did it work
List every change made in this account in the last [14] days, with the date and who or what made it. For each change, show the metric it was supposed to move, before and after, and say plainly: worked, hurt, or too early to tell.
Needs account access.
This is the prompt that catches your own mistakes, and the automated ones. "Too early to tell" appearing in the answer is a feature — an assistant willing to say it is an assistant not inventing verdicts.
5. Campaigns starved by budget
Which campaigns hit their daily budget limit in the last [30] days, on how many of those days, and what does impression share lost to budget say I left on the table? Put the campaigns worth funding first, and say what you would take the money from.
Needs account access.
A good answer pairs every "give this more" with a named source of the money. Reallocation advice that only has a destination is half a recommendation.
6. Diagnose a drop
My [cost per conversion | conversion volume] has moved about [30]% since [date]. Using this account's data, test the likely causes one by one: recent changes, auction pressure, traffic mix, seasonality checked year over year, and whether tracking itself broke. For each cause say: supported, contradicted, or the data cannot answer it — and quote the numbers that made you say so. Finish with the single most likely cause and the one thing to check next.
Needs account access.
The quoted numbers are the test. A checklist of possible causes is what you get from any assistant with no data; verdicts with your figures beside them are what the connector is for.
7. Tracking before anything else
Audit my conversion tracking before I optimise anything. Report every conversion action with its category, Primary or Secondary status, counting setting and attribution window. Flag duplicates, actions with zero recorded conversions in [30] days, and any gap between clicks and conversions that looks like measurement failure. Rank the problems by how much each distorts Smart Bidding today.
Needs account access.
Run it first, whatever brought you to this page. Every optimisation downstream of broken tracking optimises toward the wrong number — the ranking tells you which broken setting costs the most.
Search terms, in two depths
8. Live triage into negatives
Pull my search terms for the last [30] days and sort every term into: ready to buy, comparing, researching, wrong product or audience, junk. Give spend, clicks and conversions per bucket. Then propose negative keywords from the last two buckets, grouped by theme, each with campaign or ad group level and one line of reasoning I can disagree with. Propose only — add nothing.
Needs account access.
Read the proposal before approving it: that reading is the only thing between a negative list and a blocked money keyword. Five legible buckets is the mark of a good answer; fifty micro-groups means it dodged the judgement calls.
9. The long paste
I am pasting my full search-terms report below — [90] days, every row. Read all of it before answering. 1. the ten terms wasting the most money, with spend 2. patterns across the waste: themes, not individual terms 3. converting terms that deserve their own exact-match keyword 4. the negative keyword list you would build from this file Note where the export's columns limit what you can conclude. [paste the whole export]
Works in any chat.
This is the template for accounts that are not connected yet, and it leans on the thing Claude is unusually good at: a 90-day export is thousands of rows, and it fits in one paste. Two honest limits — the snapshot is stale the day after you export it, and Claude can propose the negative list but cannot apply it. When both limits start costing you time, the connector removes them.
Copy and pages
10. An ad written under oath
Here is my landing page: [URL]. Write a responsive search ad for the ad group [name] with keywords [list]: 15 headlines under 30 characters, 4 descriptions under 90, primary keyword natural in at least 3 headlines. Rule: every claim must be backed by a line on that page. After the ad, list each claim with the exact sentence backing it. A claim with no backing gets flagged, not softened.
Works in any chat.
The claim-to-quote list makes the whole ad reviewable in a minute, and it is where message-match failures — a top cause of expensive clicks — get caught before they ship. Count distinct angles in the headlines, not headlines.
11. The objections your page never answers
Read [URL] as a sceptical buyer who has just clicked an ad and is deciding in 30 seconds. List every objection the page raises but never answers — price, risk, effort, trust, "is this for me" — and for each one say where on the page the answer belongs.
Works in any chat.
Usually the cheapest win on this page: every unanswered objection is a conversion you already paid for and lost. The location half of the answer is what turns the list into a to-do.
Change things, carefully
12. Reallocate without resetting learning
Using the last [60] days, propose budget changes for the next [30]. Show every campaign: budget, spend, conversions, cost per conversion, and whether it was budget-limited. Rules: never cut a budget-limited campaign without saying so, no single change over [20]%, total stays at [amount] per day, and flag anything the change would push back into learning. Propose in a table with reasoning and an order of operations. Change nothing.
Needs account access.
The two classic reallocation errors are moving money away from a campaign that never had room to prove itself, and moving it all at once — the budget-limited rule blocks the first, the 20% cap blocks the second.
13. A campaign built paused
Using the business context in this project, build a Search campaign: goal [conversion action], budget [amount]/day, geography [list]. 4 to 6 tight ad groups; 8 to 15 keywords each with match types; one responsive search ad per group; a themed starting negative list; your bidding strategy choice argued from my actual conversion volume. Show me the full structure first and wait for my go-ahead. Create it paused. Say what you cannot know until it runs.
Needs account access to create it; the structure alone works in any chat.
"Using the business context in this project" is prompt 1 paying rent — no re-briefing. If the reply does not say "paused" in writing, open the account and check before anything else.
Keep it running
14. The weekly review you actually repeat
Weekly review, one screen: last week vs the week before by campaign (spend, clicks, conversions, cost per conversion, change); unusual budget pacing; search terms over [20] spend with zero conversions; search terms appearing for the first time; changes made and by what; then the one thing to fix first and why that one.
Needs account access.
Claude has no scheduler — unlike ChatGPT's Tasks, nothing here runs on its own. The workaround is structural: keep this prompt pinned in the same project as prompt 1 and let a Monday calendar reminder do the scheduling. The value is the comparison arriving weekly, not the prompt being clever.
15. The report as an artifact
Write a performance summary for [client or stakeholder] covering [period], as an artifact I can export and send. Audience: [technical | non-technical]; they care about [cost per lead | revenue] above all. Three plain sentences on what happened; the numbers with the previous period beside each; what we changed and what each change did; what is not working, said plainly; next month's plan. Never call a drop an opportunity.
Works in any chat, far better connected — then the numbers are your own.
Asking for an artifact is the point: the summary arrives as a standalone document rather than chat prose, editable in place and ready to send. The "said plainly" clause is there because reporting drifts toward flattery on its own — the summary that names what is broken is the one that keeps the client.
The rules doing the work
Three clauses repeat across all fifteen, and they are the transferable part. A stated window and a comparison — "last 60 days", "versus the previous period" — because an unstated window makes two runs of the same prompt incomparable. A refusal instruction — never estimate volume or CPC — because Claude, like every assistant, invents those numbers fluently when allowed to. An approval clause — propose only, create it paused — which is also how the connection itself behaves: writes arrive as proposals and campaigns are born paused, as the documentation explains. Everything else is your judgement, and prompt 1 is where it lives.