All posts

Claude Skills for marketing analytics: the ones worth building.

Four narrow Skills for the marketer who owns the reporting. What each one triggers on, what it knows, what it refuses, and the one refusal that makes Claude read like a senior analyst instead of a junior one.

Analytics is the discipline where a generic prompt hurts you most. Ask Claude to "summarise this data" with no rules and it will do the thing every junior analyst does: narrate the dashboard back to you, call every wiggle a trend, and bury the one number that mattered under nine that did not. The fix is not a better prompt you paste each Monday. The fix is a set of Skills that load on their own and enforce the rules whether you remembered them or not.

This post covers the four Skills a marketer doing analytics and reporting should build, what goes in each, and a worked SKILL.md for the most important one. If you have not built a Skill before, the 20-minute build walkthrough has the full mechanics. This is the analytics-specific version: which Skills earn their slot, and why.

The four Skills worth building

Three to five narrow Skills beat one bloated "Analytics" Skill every time. For reporting work, these four cover the job. Each one earns its trigger and does one thing.

The Skill that does the most work: a worked SKILL.md

Of the four, the Performance Readout Skill is the one to build first. It is the one that changes how every report reads, and it pairs with a prompt you can run for free today (the free Performance Readout prompt). Here is the whole file.

---
name: performance-readout
description: Use this Skill whenever the user is reading,
  summarising, or interpreting marketing performance data
  (a pasted CSV, a dashboard export, GA4 numbers, an ad
  account, a monthly or weekly metrics report) and wants a
  readout rather than a description. Triggers on "summarise
  this," "how did we do," "read this report," "what does
  this data say," "write the monthly readout," or any pasted
  table of marketing metrics. Enforces one decision sentence
  plus exactly three findings, and refuses to narrate.
---

# Performance Readout

## What this Skill knows

A readout is not a description of the dashboard. It is a
short answer to one question: what should we do about this?

Every readout has the same shape:
1. One decision sentence first. The single thing the reader
   should do, change, or keep doing, stated plainly.
2. Exactly three findings, ranked by what they change. The
   most decision-relevant finding goes first.
3. One "what we are watching" line for the thing that is
   not actionable yet but might be next month.

Metric definitions live in metric-definitions.md (this
folder). Use those, not textbook ones.

## What this Skill does

When loaded, on any performance data:
- Lead with the decision sentence. Never bury it.
- Give three findings, no more. If a fourth feels
  important, it replaces a weaker one. It does not get added.
- For each finding, name the number, the change, and the
  "so what." A finding without a "so what" is a stat, cut it.
- Flag any number that looks like a tracking artefact rather
  than a real move (a sudden zero, a doubled count, a gap).

## What this Skill refuses

- Do not narrate the dashboard. No "sessions rose then fell
  then rose." Movement is not a finding. A decision is.
- Do not pad with an executive-summary paragraph. The
  decision sentence is the summary.
- Do not report more than three findings. The discipline is
  the point. Ranking forces a real call about what matters.
- Do not declare a trend from two data points. If the data
  is two periods, say "two points, not a trend" and stop.
  Defer to the refuse-the-trend Skill if it is loaded.
- Do not invent a cause for a move. "Up 12%" is a fact.
  "Up 12% because of the campaign" is a claim that needs
  the data to support it. If it does not, say "cause unclear."

The refusal that matters most: refuse the trend

This is the move that separates a senior analyst from a junior one, and it is worth its own Skill. A junior analyst sees two months of rising signups and writes "signups are trending up." A senior analyst sees the same two points and says: that is two data points. It could be noise. It could be seasonality. It could be the one big customer who signed twelve seats. Come back when there are enough points to tell.

Claude defaults to the junior move. It is agreeable, and "the trend is up" is the agreeable thing to say. A refuse-the-trend Skill makes it do the senior thing instead: refuse to call a trend until the data earns it. The instruction is short.

## What this Skill refuses

- Refuse to call any movement a trend from fewer than the
  threshold the user sets (default: at least 5 periods, and
  a move larger than the normal period-to-period swing).
- Two points is a line between two numbers. Say so, then
  stop. Do not extend the line into a forecast.
- If the data is too short to judge, say exactly that:
  "Not enough periods to call this. Here is what it looks
  like so far, and here is what would confirm it."
- Never attribute a single-period move to a single cause
  without evidence in the data that rules out the others.

The reason this works as a Skill and not a prompt: it has to fire on every report, including the ones where you forgot to ask for it. The week you are rushing to get the board deck out is exactly the week you would skip "and please do not over-read the data," and exactly the week an over-read finding does real damage. A Skill does not skip. That is the whole value of it being infrastructure instead of a habit.

How the four fit together

Installed together, the four Skills compose. You paste a CSV. Performance Readout fires and shapes the output into a decision plus three findings. Metric Definitions fires and makes sure "qualified lead" means your qualified lead. Refuse The Trend fires the moment Claude is tempted to call a two-point move a trend. Readout Format fires when you ask for the actual deliverable and locks the shape. None of them is big. Each does one thing. Together they turn Claude from a dashboard narrator into the analyst you wish you had on the team.

Build them in that order, too. Performance Readout first, because it changes every report immediately. Refuse The Trend second, because it is the lowest-effort insurance against the most expensive mistake: telling the board something is working when it is just noise. Metric Definitions third, once you have written your house definitions down somewhere a Skill can point at. Readout Format last, because the format only matters once the thinking underneath it is right. Most marketers stop after the first two and already feel the difference.

One note on the split between them. Refuse The Trend and Performance Readout overlap on purpose. The readout Skill carries a short version of the refusal so it works on its own, and the dedicated Skill carries the full threshold logic for the weeks when the data really is borderline. That redundancy is fine. A refusal you state twice is a refusal that survives the week you were not paying attention.

If you run reporting for a team rather than just yourself, the same Skills are how you make the weekly readout reproducible across people. That is the setup in the weekly marketing readout workflow: the Skills carry the standard so the readout reads the same whoever ran it.

The honest caveat

A Skill cannot fix bad data. If your conversion tracking is broken, a Performance Readout Skill will give you a beautifully structured readout of broken numbers. The Skill enforces how the data is read, not whether the data is right. Get the tracking and the definitions sorted first. Then the Skills make sure nobody ever has to remember the rules again.