Your First Prompts

You are connected. Here are the questions that work right away. All read-only, nothing gets changed.

Every prompt here has been tested on a real account, none of this is theory. The examples in each step show you the shape of the answer to expect: your numbers will differ, the structure and the interpretation won't.

Whether you use Claude, ChatGPT, or your own agent, the flow is the same. We say "your agent" here.

A quick word first: which client?

A workspace is one client with their connected data sources. Two cases:

• You have only one workspace: Nothing to do, your agent picks it automatically. Step 1 is still worth it as a source check, step 2 you can skip.
• You have several (agency): Say which client you mean in the first prompt. After that it stays in the conversation, you don't repeat it. New chat, new client: name it once. And if you forget, the server stops and shows you the options (see the box under step 2).

What you need

A connected connector. If you have not connected yet, start with Connect your marketing accounts to Claude. Steps 1 and 2 only need the connector. After that, each step depends on its source: steps 3 and 4 need GA4, step 5 the client's CMS, step 6 Microsoft Clarity. Steps 7 and 8 need Search Console, step 9 Search Console plus the CMS, steps 10 to 13 a Google Ads account. Step 14 only bundles what is already in the chat and needs no new source. If a source is missing, skip those steps.

The prompts

1. Get the overview

Right at the start, especially as an agency: you have one workspace per client. Get the list and, from it, the shortname you will use to address each client.

Use the marketing-mcp: Which workspaces do I have?
Show me shortname, client name, connected data sources, and whether I have write access per workspace, as a table.

Here is what you get back: your client list with shortnames, sources, and write access. How your agent presents it is up to the agent, which is why the desired format ("as a table") is part of the prompt. That is how every prompt here works: if you want a table, say table. For us the list looks roughly like this (other clients anonymized):

• Our own website (honeyfield-gmbh): all eleven sources (Google Ads, GA4, Search Console, GTM, Business Profile, Clarity, DataForSEO, Meta Ads, LinkedIn Ads, Strapi, WordPress)
• Client (anonymized, client-a): GA4, Search Console, Business Profile, DataForSEO
• Client (anonymized, client-b): GA4, Search Console, GTM, Clarity, DataForSEO, Strapi
• and one more per client

The shortname in parentheses (e.g. client-a) is what you need in step 2. If you have only one workspace, treat this step as a pure source check and jump to step 3.

2. Set the client for this chat

Only needed if you have several clients. One prompt, three effects: it puts the client into the chat, confirms it is the right one, and makes it visible in every answer from now on.

In this chat we are working with the workspace `client-a`.
Confirm briefly: Which data sources are connected there, and do I have write access?
From now on, add to every answer which client the numbers belong to.

From here on the client is part of the conversation, your agent passes it along with every request, you don't repeat it. The source answer is your check that the right client is meant before the first numbers flow. This only holds for this chat, a new chat means naming the client again. For us, with our own shortname in the prompt, the agent confirms all eleven sources and write access.

Why the last line is part of the prompt: The guardrail below only kicks in when no client is named. If the wrong one sits in the conversation, the server happily returns wrong numbers. That one line per answer makes exactly this case visible. Costs one line, worth the price with multiple clients.

When the follow-up question comes. If you name no client and have several, the first data request usually stops and shows a message along the lines of:

"This account has access to several workspaces, please provide the 'workspace' parameter. Available: client-a (AgencyX), client-b (AgencyX)."

That is not a glitch, that is the guardrail: better to stop once than to report numbers for the wrong client. Just answer in the chat: "Use client-a." The request is retried, nothing is lost. Some agents ask on their own before fetching anything, the answer is the same: name the shortname and continue. (The 'workspace' parameter means exactly your shortname.) If you mistype a shortname, you get something like "Workspace '…' not found. Available: …", also harmless, just name the right one.

3. How many visitors, and where is the trend heading?

The first number that matters. "The website" from here on automatically means the one of the client from step 2, who is part of the conversation.

How many visitors did the website have in the last 28 days? Show me the trend.

Here is what you get back: active users and sessions for the last 28 days, the comparison to the previous period, and a trend chart, for example: 420 active users, 492 sessions, up 57%. Your agent warns you on its own that the last three days are GA4 processing lag, so still incomplete. And it interprets instead of just counting: a one-time jump in the curve does not automatically count as "growth".

Your benefit: the most important metric plus interpretation in one answer, without opening GA4.

4. Where does the traffic come from?

Channels instead of gut feeling. Where the sessions really come from. Name a concrete window ("last 7 days"), not "this week": that way your agent cleanly excludes the unfinished GA4 days.

Where did the website traffic of the last 7 days come from? Group by channel.

Here is what you get back: the sessions by channel, for example:

• Paid Search: 69% of sessions
• Direct: 14%
• Organic Search: 8%
• Referral: 8%

On top, your agent adds the engagement rate per channel, the share of visits that stay or do something. So you see where people come from and which channel actually keeps them.

Your benefit: One question tells you whether your traffic is earned or bought, something you would otherwise combine two GA4 reports for.

5. Which articles are live?

A look into the CMS without logging in.

Show me the titles of the last 10 blog articles on the website (Strapi).

Here is what you get back: the titles of the last ten articles with date and status, straight from the CMS, plus the total number of entries.

Two CMS? On the Honeyfield website we have WordPress and Strapi connected, WordPress only as a test instance though. Without the (Strapi) in the prompt your agent guesses and may pick the wrong source. With two CMS, name the source. If your client has only one (the normal case), you can drop it.

Your benefit: check the editorial state without a CMS login, even for client websites where you have no account of your own.

6. Where do visitors get stuck?

Frustration signals from real sessions, not from assumptions.

Where do visitors get stuck on the website? Show me rage clicks, dead clicks, and scroll depth.

Here is what you get back: the frustration signals from Microsoft Clarity, for example: rage clicks 0 · dead clicks 22% · avg. scroll depth 41%. Plus the data basis, because Clarity only covers the last 3 days. Your agent tells you honestly whether the sample is big enough for conclusions, instead of turning 20 sessions into truths.

Your benefit: real user signals instead of assumptions, including a warning when the data basis is thin.

7. Your top Google search queries

From here on you need Search Console. The words people use to find the client.

What are the top Google search queries of the last 28 days?

Here is what you get back: the top queries with clicks, impressions, CTR, and position. The interesting part is the patterns your agent adds: how many clicks come from your own company name (brand) and how many from real topic searches. Or keywords that rank high yet barely get clicked, up to position 1 with zero clicks. The ranking is already there, only the title and description in the search result (the snippet) don't convince anyone yet.

Your benefit: Instead of a keyword table you get the cheapest SEO lever named: start where visibility already exists and only the click is missing.

8. Which pages bring the clicks?

Same source, different angle: pages instead of queries.

Which pages get the most clicks from Google Search?

Here is what you get back: the pages with the most clicks from Google Search, including the outliers, for example a single article with 1,435 impressions and one single click (CTR 0.07%). Exactly these pages are the biggest single lever: lots of visibility, no return. By the way, this page view usually counts more clicks than the query view from step 7, that is normal: there, Google omits anonymized queries.

Your benefit: The outliers hiding in almost every account, found with one question instead of you walking through all pages by hand.

9. Which articles rank, and which don't?

The prompt no dashboard can do: your agent combines two sources in one question. It pulls the articles from the CMS and matches each one against its clicks from Google Search. It matters that you name the CMS source and explicitly demand the cross-check. Otherwise the agent takes the shortcut and only filters Search Console, and then every article without impressions is missing.

Pull the blog articles from the CMS (Strapi) and match each one against its organic clicks from Google Search.
Which ones get clicks, which get none at all?

Here is what you get back: every article from the CMS with its organic clicks, in three groups: articles that get clicks. Articles that show up on Google but don't get clicked. And articles without a single impression, often because they are not indexed yet, say a freshly published batch of new posts. Exactly this third group is invisible to a pure Search Console view, because there only things with impressions exist. The cross-check with the CMS sees it.

Your benefit: the complete content balance from two sources in one answer, including the articles every pure SEO tool overlooks.

10. How did the Google Ads campaigns do?

From here on you need a Google Ads account. Which client has one is in your list from step 1. The last 7 days at a glance.

How did the Google Ads campaigns perform in the last 7 days?

Here is what you get back: all campaigns with status, impressions, clicks, cost, CTR, CPC, conversions, and CPA. Your agent double-checks instead of just reporting: a CTR above 40% is not a success but a warning sign (barely any delivery at tiny volume). And it asks what a "conversion" measures in your setup, because form submissions are not leads yet.

Your benefit: a campaign overview that flags the implausible instead of selling it to you as success.

11. Are there any anomalies in the campaigns?

You don't have to search yourself. Your agent scans the last days for cost spikes, dropped conversions, stopped delivery, and CTR drops. Unlike GA4 in step 4, "this week" is fine here: Ads numbers come in almost in real time, and the scan compares the most recent days anyway.

Are there any anomalies in the Google Ads campaigns this week?

Here is what you get back: the anomalies of the last days, for example "delivery at €0 for two days" or "CTR halved versus last week". The most valuable find is the silent failure: a campaign set to ENABLED that still delivers nothing, say because of payment or disapproved ads. Dashboards rarely catch this, because nothing turns red. It just goes quiet.

Honestly: on a workspace without a Google Ads account, the prompt tells you exactly that instead of inventing something.

Your benefit: You learn about dead campaigns before they stay unnoticed for days, with one question instead of a daily control check.

12. Is the budget on track?

Pacing shows whether you will overspend or underspend by month's end.

How is the budget pacing of the campaigns this month?

Here is what you get back: daily budget, spend so far, pacing in percent, and the projection to month's end. More important than the number is the interpretation: a pacing of 80% can mean "all calm", or overspend first and standstill after, the average smooths out both. That is why your agent checks the daily curve before waving through Google's reflex ("raise the budget").

Your benefit: a budget statement that knows the trajectory, instead of an average that irons out problems.

13. Which search terms burn money?

The prompt that makes all of this worth it. Money out, conversions zero.

Which search terms cost money but bring no conversions?

Here is what you get back: the search terms with cost, clicks, and conversions, sorted by wasted budget, for example:

• "free [your service]": 62 clicks · €174 · 0 real leads → search intent doesn't match
• "[off-topic term]": 23 clicks · €57 · 0 conversions → different industry

Clusters like these usually come from broad match (Google also serves your ad on loosely related searches) and go unnoticed for months without this report.

Your benefit: One answer shows you which part of the budget goes to searches that never buy, and delivers the ready-made candidate list for negative keywords with it.

This is exactly the moment where connecting pays off. A skill, a ready-made workflow for your agent, takes it from here and adds the negatives (negative keywords), after your approval.

14. Turn everything into a report

The finale. Up to here, every answer was one prompt, one question. Now your agent bundles everything from the chat into one report. This only works because all analyses live in the same chat (one chat, one client), it has the whole case in memory.

Create a report on the situation from all of your analyses. Use a light design.

Here is what you get back: a standalone report on everything analyzed in this chat. In Claude as a finished HTML page, other agents deliver it in their own form. Your agent picks the thesis that connects the analyses itself, builds its own visualizations for it, and ends with the honest limits of the data, not just the pretty numbers.

"Use a light design" only controls the look. Say "dark", "newspaper style", or nothing at all, and your agent decides.

Your benefit: a shareable report without copy-paste from five tools, created from a single sentence.

At the end: the second client

The whole model in one move: open the next client. Open a new chat and start with:

In this chat we are working with `client-b`. Which data sources are connected?

(A different shortname from your list.)

The old chat knows nothing about client-b, the new one nothing about client-a, so nothing mixes in the conversation and every request cleanly belongs to one client. That is the whole model: one chat, one client, named in the first prompt. From here, every additional client is just another new chat.

Switching or comparing two clients mid-chat works too, the server doesn't forbid it. But every switch raises the risk of mix-ups, because two clients are then part of the conversation. A recommendation, not a rule: one chat per client, and if you do switch, name the new client explicitly afterwards.

One prompt is just the start

All prompts here only read. Nothing gets changed, not a cent gets spent. That is deliberate, and it is only one half. The same connector can also write: publish a blog article straight into the client's CMS, create and pause campaigns, adjust budgets, answer reviews. None of it happens unasked, you approve, then it writes. Almost every prompt here was one question, one answer. The report at the end bundled them, but it, too, only read. A skill goes one step further and turns the analysis into a workflow that finds the expensive search terms, proposes the right negatives, and adds them after your approval.

Doing this daily? One project per client. If you often work for the same clients, skip the "name the client again". Create one project per client (both Claude and ChatGPT have projects) and put one line into its instructions: "Use the marketing-mcp, workspace client-a." Every chat in that project then works in the right workspace without naming it again. And since only one client lives per project, the risk of mix-ups drops sharply. The client folder stays on the desk, cleanly separated per project.

What's next: The next step is your first write action, with approval. In Find Wasted Spend and Set Negatives your agent proposes the negatives, you confirm, and only then are they applied.