Write a Sponsored Ads SOP, then turn it into your first Claude skill

What makes a Sponsored Ads SOP something a new hire can follow?

It names the data, the rule and the person who confirms, on every step. "Review search terms weekly" is a reminder. "Pull 30 days of search terms, flag any that spent more than [X] with zero orders, get a senior to approve the negatives" is an SOP. New hires rarely trip on judgment first. They trip on gaps: which window, which threshold, who says yes.

Most teams keep that routine in one person's head and a half-finished doc. The same document can become your team's first skill. Xnurta MCP connects Claude, ChatGPT and other MCP clients to your Xnurta data, and a skill teaches the assistant a routine on top of that connection. If you have not connected yet, start with Connect Claude or ChatGPT to Xnurta MCP.

What does a weekly search term SOP look like?

Seven steps, and five of them only read. The template below covers a Monday search term review for one ad account. Swap every bracket for your team's own number, because the thresholds are yours, not ours.

Step 1 sits first on purpose. A new hire who negates a term a teammate just promoted has undone someone else's work, and the log is where that shows up. Steps 2 to 5 are the analysis. Step 6 is the only one that writes, and it needs write access on your account. Step 7 is boring and it is the one teams skip.

How do you run the SOP with an assistant before you write a skill?

Run it by hand first, in one chat, one step at a time. A skill built from a messy routine automates the mess. Run the SOP once, note every place the assistant needed a correction (wrong window, missing threshold, unclear output), and fix the SOP before you do anything else.

Three habits make the run cleaner. Name the object, meaning the ad account and campaign, right in the question. Give a time window and a comparison. Spell out the output you want, whether that is a table, a top-N list or a one-line takeaway.

In ad account [account name], review search terms for the last [30] days. Show CTR, CVR, ACOS and orders. List terms that spent more than [X] with no orders and have not been negated. Then list terms that spent [X] to [Y], produced orders, and were negated, so I can check for wrong negatives. Then list terms from auto targeting with CVR above [X]% and more than [two] orders that belong in manual campaigns. Change nothing.

From the lists above, show every negative keyword and every manual-campaign addition you would make: campaign, ad group, keyword, match type and bid. Do not submit anything. Wait for my confirmation.

Which official skills show you how to write one?

Weekly Ads Report, Monthly Ads Report, Ads Structure Analysis and Product Diagnosis are four worked examples of a skill that someone already got right. Each one fixes the data it inspects, the order it works in and the shape of its answer. Read them the way a new hire reads your SOP.

Notice what repeats. A trigger phrase a person actually says. A fixed list of things to check. A fixed output, so Tuesday's report looks like last Tuesday's. Borrow all three.

To try one, name the skill you want, such as "I want the weekly ads report", and the assistant guides you through installing it. Connecting never installs skills on its own. In a new chat, ask the assistant to follow the instructions in the Xnurta skills repository. Search Term Traffic Diagnosis is also on the official list, so install it and compare it with your SOP. More on the full set in How an SOP becomes something an AI assistant can run.

What are the three parts of a skill?

Scenario, rules and output. Those are the three steps the assistant takes you through when you build one, and they map onto your SOP almost one to one.

• Scenario. When the skill applies and what the person says to start it. For the SOP above: "Every Monday I review search terms for one ad account at a time, and I name the account."
• Rules. The thresholds, the order of steps and the lines it never crosses. This is where your brackets go. A good rule is also a refusal: "Never archive anything. Pause or negate instead."
• Output. The shape of the answer. "One table per flag type with campaign, term, spend, orders, CVR and ACOS, then a one-line takeaway, then the list of proposed changes."

If a step in your SOP cannot be sorted into one of the three, it is probably a habit, not a rule. Cut it or make it concrete.

How do you build the skill with Xnurta MCP?

In a client that supports skills, tell the assistant "I want to build a Skill." It guides you through three steps: describe the scenario, set the rules and define the output. Use your SOP's content for each of the three and answer what it asks.

I want to build a Skill called Weekly Search Term Cleanup.

Scenario: Every Monday I review search terms for one ad account at a time. I will name the ad account.

Rules: Check the operation log for the last [14] days first, so we do not undo a teammate's change. Use the last [30] days of search terms. Flag terms that spent more than [X] with no orders and are not negated. Flag terms that spent [X] to [Y], produced orders and were negated. From auto targeting, flag terms with CVR above [X]% and more than [two] orders as candidates for manual campaigns. Never archive anything. Before any change, list every object and exact value and wait for my confirmation.

Output: One table per flag type with campaign, search term, spend, orders, CVR and ACOS. Then a one-line takeaway, then the list of proposed changes.

When you have drafted it, test it on [account name] with my real data.

Run Weekly Search Term Cleanup on [account name] for the last [30] days. Compare its flags with the list I made by hand this week and tell me every difference.

How do you test the skill and share it with the team?

Test it on real data, then compare the result with a run you did by hand. The assistant can test-run the skill on your real data while you build it, and you save it once you are satisfied. The second prompt above gives you a clean way to judge it: every difference is either a gap in the skill or a mistake in your hand run.

Then you choose who gets it. You can keep a saved skill private or share it within your organization, and teammates can install what you share. Keep it private while you iterate. Share it when two people run it on the same account and get the same flags. Keep the SOP as the page your team reads and the skill as the way the assistant follows it.

What stays in your control when a skill touches your ads?

The same controls as without a skill. Reading performance, settings and the operation log changes nothing in your account. Edits need write access, and each kind of write has its own permission scope, so a connection that can edit ads cannot necessarily create campaigns.

Anything involving money, bids or archiving returns a preview first: which objects, and from what value to what value. Nothing is submitted until you confirm. Archiving cannot be undone, so pause a campaign if you only want it to stop. For how the read and write sides differ, see Read-only vs read-write MCP for Amazon advertisers.