Ten Skills Or One Agent

The short version

Teams building AI into marketing operations face the same design decision repeatedly: build another small skill for this task, or hand the whole workflow to something that owns it end to end.

This post gives a checklist for making that call, and a way to recognize when a skill library has stopped paying off.

The Decision In One Line

Build a skill when the task is a step. Hand it to an agent when the work is a loop.

The difficulty is that most marketing work looks like a set of steps until you try to automate it, and then turns out to be a loop.

The Checklist

Score the workflow. Each yes moves it toward an agent.

Question	Yes means
Does the output of one part become the input of another?	The handoff is work, and a skill will not do it
Does doing it well require knowing what happened last time?	Memory is required, and skills are stateless
Does the right next action depend on what you find?	The sequence is not fixed, so it cannot be encoded
Would two competent people sequence it differently?	Judgment is involved, not procedure
Does it run on a cadence rather than on a trigger?	Something must decide when, not just what
Does skipping a step sometimes turn out to be correct?	The path is conditional

Zero to two yeses is a skill. Four or more is a loop, and building it as skills will produce the sprawl pattern instead of the workflow.

Worked Examples

Generate five headline variants from a brief. No handoff, no memory, fixed sequence, one correct shape of output. This is a skill. Building it as an agent adds cost and unpredictability for nothing.

Summarize last week's account performance. Reads data, produces a report, sequence is fixed. Still a skill, even though the input is large.

Find out why CPA rose and do something about it. Diagnosis determines the response. The response depends on what the diagnosis found. Doing it well requires knowing what was already tried. Two analysts would sequence it differently. This is a loop, and cutting it into skills leaves a person doing the actual work between them.

Keep creative fresh across three platforms. Requires knowing what is running, what is fatiguing, what was tried, and what to produce next, with the answer changing weekly. A loop.

The Cost Nobody Counts

Each skill has a maintenance and coordination cost that does not appear when you build it.

Someone must remember it exists and when to use it.
It breaks silently when the underlying data or platform changes.
Its output format must stay compatible with whatever consumes it next, which is usually a human.
Its usefulness depends on being invoked at the right moment, which is a decision nobody automated.

Ten skills means ten of these. The coordination cost is why the tenth skill delivers less than the second, and why teams plateau without understanding why.

When A Skill Library Is The Right Answer

Skills are correct when the tasks are genuinely independent, the people using them know when to reach for which, and no output needs to travel automatically from one to another. A creative team with a set of well-scoped generation helpers is a reasonable design, and turning it into an agent would be a downgrade.

The failure is not having skills. It is having skills where a loop was required.

Common Mistakes

Building the eleventh skill to solve a problem caused by the previous ten not connecting.
Treating "we can automate each step" as evidence that steps are the right decomposition.
Measuring progress by skills shipped rather than loops closed.
Handing a genuinely fixed procedure to an agent, which costs predictability for no gain.

Practical Rule

Run the checklist before building. If the workflow scores four or more, do not start by building the first skill, because the skills are not the hard part and building them first makes the loop harder to see.