SEO SOP: Build a Claude Skill That Does Your Pre-Work

Before a strategist can analyse anything, an hour or two goes into pulling numbers out of several systems and lining them up in a spreadsheet. That assembly work is what an AI skill should absorb. Everything after it stays human, deliberately. The play is Erika Braeger‘s, Manager of Organic Growth Strategy at the B2B SaaS content agency Ten Speed, and she laid out her method on Unscripted SEO with Jeremy Rivera. This SOP turns it into something you can run on one recurring deliverable this week.

Objective

Take one analysis your team runs over and over, find the assembly work that happens before the thinking starts, and build a single AI skill that produces that starting artifact on demand. The strategist stops opening four tabs to reconstruct where an account stands and instead opens a finished input. The analysis, the narrative, the strategic decision and the client relationship all stay with the human, on purpose. Budget a couple of days to build the first one, plus one full delivery cycle running it beside the human before you trust it.

Key Steps

1. Time the pre-work before you automate anything. For two weeks, have the team note on each recurring deliverable where the time goes before analysis starts. You are looking for assembly, not judgement: numbers pulled out of several systems and lined up somewhere. Braeger’s description of the problem is the test to apply. “That pre-work takes time, it takes brain power before you can actually begin the analysis, before you can make a decision, before you can craft a narrative for the client, and before you can help the client execute.” If a task fails that test, it is not pre-work, it is the work.
2. Pick the most boring repeated job on the list, not the most impressive one. Her own worked example is deliberately mundane: conversations scattered across systems, gathered and summarised, so nobody has to jump from place to place to reconstruct where an account stands. Pick the equivalent in your shop. A monthly performance roll-up, a pre-call account state, a competitor position refresh. Frequency beats sophistication, because a skill only pays back on repetition.
3. Write down what the skill will not do, first. This is the step teams skip and it is the one that decides whether the play works. Braeger draws the line explicitly: “I don’t want to take away the thinking part. I want to allow the strategist to arrive at the thinking part with more ease, without burning energy, on grabbing data, organizing it out of multiple systems, organizing it in the spreadsheet.” Put the exclusions in writing before anyone builds. No recommendations, no narrative, no client-facing conclusions, no strategic decision.
4. Specify the output artifact before you specify the process. Describe exactly what the strategist should have open when they sit down. The columns, their order, the date range, how a summary is shaped, what a missing value looks like. Then build backwards from that artifact. A skill with a vague brief produces something a strategist has to reorganise, which is the same pre-work wearing a different hat. (inferred: she specifies the outcome she wants rather than a build method.)
5. Name every source system and confirm access before writing a line. List each place the data actually lives, then check the skill can reach every one. This is where most builds stall, and it is an access problem rather than an intelligence problem. Braeger’s own list of places worth watching has grown well past the analytics stack: “So now I need to pay attention to Reddit and I need to pay attention to G2 and I need to pay attention to YouTube.” If one source cannot be reached, decide up front whether the skill ships without it or waits.
6. Run the skill beside the human for one full cycle. Have the strategist assemble the artifact by hand as usual, then compare it to what the skill produced. You are looking for what it silently dropped, not only for what it got right. Close those gaps before anyone relies on it. (inferred: the parallel run is standard practice for a handover like this, and follows from the QA insistence below.)
7. Keep a QA pass on every run, permanently. This is not a launch checklist you retire once it settles. Braeger is direct that the check does not go away, even after a lot of skill-building: “I still need to run a thorough QA.” Asked whether we will be able to hand more over as these systems improve, her answer was “I’m hesitant.” Assign the QA to the strategist who uses the output, because they are the one who will notice a number that looks wrong.
8. Measure the returned hours by where they went, not by how many there were. The point is not a saving on a timesheet, it is what the reclaimed brain space gets spent on. Braeger’s own list is the scoreboard: running the analysis, crafting the narrative, making the strategic decision, helping the client execute, and growing the people on the team. If the hours quietly refill with more accounts per strategist, you bought throughput instead of quality and the play did not land.

Cautionary Notes

• The moment the skill drafts the recommendation, you have automated the wrong half. Clients pay for the analysis and the narrative. Absorbing the assembly and keeping the thinking is the entire design, and it is easy to reverse by accident, because a model will happily offer a conclusion at the end of a data pull.
• A QA pass longer than the pre-work it replaced is a net loss. If checking the output takes as long as building the artifact by hand did, the skill is either too ambitious or too vaguely scoped. Narrow it until the check is fast.
• Do not build it for a job that runs twice a year. The return comes from repetition. Anything infrequent enough that the source systems change between runs will rot before it pays back.
• Junior strategists lose the reps. Braeger’s management method is coaching people through gut checks until they trust their own instincts. Someone who has never assembled the data by hand has a weaker sense of when a number looks wrong, so keep newer strategists doing the manual pass for a while before handing them the skill. (inferred: she did not connect these two threads on the episode, but they collide in a real team.)
• Access is the blocker, not capability. Read permissions, seats, API keys and export limits kill more of these builds than prompt quality does. Sort that out before you scope the work.
• Some of the pre-work has no source you control. Tracking mentions across community platforms and video is the current gap, and it is Braeger’s own unsolved problem rather than a solved step. Scope around it instead of pretending the skill covers it.

Tips for Efficiency

• Start with the deliverable your team complains about most. The complaint is free research, and adoption is already handled.
• One skill per artifact. A single skill that produces five different outputs is five vague skills sharing one QA problem.
• Make the output identical every time. Boring and predictable is the feature, because a strategist should be able to read it without really reading it.
• Ship the first version at eighty percent and let the strategist finish the last stretch by hand. A skill in use gets better. A skill in progress does not.
• The same pattern transfers straight to new-account onboarding: product, ICP and current state assembled once, so the kickoff starts at questions instead of at data entry.
• Write the exclusion list into the skill itself, not into a document nobody reopens.

Why this one matters

Most AI adoption in agencies has been aimed at the visible, expensive part of the job. The writing, the recommendations, the deliverable a client actually reads. Braeger aimed at the opposite end, at the invisible hour nobody bills and nobody enjoys, and it is a better target precisely because nothing is lost when it goes.

She named skill-building as her biggest win of the year, and the example she reached for was not impressive at all:

“Claude can summarize conversations across your system. So instead of jumping from like place to place to place to read conversations, you can have Claude go out, grab them, summarize them, and that saves you a lot of time and effort.”

That is the shape of the whole play. Nothing clever happened. Someone stopped opening four tabs.

“It’s the pre-work that can burn a lot of time, a lot of energy. And if someone else is helping you with pre-work, whether it’s someone on the team or you’re using an AI skill in Claude, you’re able to then get back the time to invest into running your analysis, crafting the narrative, making a strategic decision and helping the client execute.”

Her closing challenge on the episode was the same idea pointed at a harder problem, keeping track of a client’s visibility and mentions across Reddit, G2 and YouTube: “So what if we’re able to take out all of this pre-work and just have everything pulled together into one place so you can start your analysis?” That one is still open. The version in this SOP is not.

Sources & Relevant Episodes

• Erika Braeger, Ten Speed – the play is hers. As Manager of Organic Growth Strategy at the B2B SaaS content agency, she builds Claude skills that remove the data-gathering and spreadsheet assembly ahead of an analysis, and stops deliberately at the thinking. Full interview: Erika Braeger on Confident Teams and What AI Actually Cites.
• Listen: Use AI to Kill the SEO Agency Pre-Work, Not the Thinking on Unscripted SEO. Watch: the full episode on YouTube. Read: Kill the Pre-Work, Not the Thinking on Substack.
• More from Erika: tenspeed.io · LinkedIn · Ten Speed’s first research report, what AI actually cites for B2B evaluation-stage prompts.
• Also on SEO Arcade: the full recap of this conversation.
• Related reading: SEO SOP: Decide the One Thing to Ship Next, which handles the decision those reclaimed hours are supposed to fund, and SEO SOP: One Episode Into Twelve Link Assets for the same build-it-once logic applied to production.
• Part of the SEO SOP library.