Open-source analytics skills for your AI agent
Your agent can get the number.
The job is knowing what to do with it.
Requesting a metric is cheap now. Sizing a move against the goal and recommending the next step still takes an analyst. aftergrid teaches your agent that work.
Works with a CSV, a notebook, or the SQL connection your agent already has. Nothing else to set up.
Conversion fell.
Someone says roll back signup.
The math is right. It points the team at the wrong fix.
analysis-review Challenge the rollback claim before the meeting.
Both channels improved.
Traffic shifted to paid.
- 10→11% Direct conversion rose.
- 4→5% Paid conversion rose.
- −3.6pp Paid grew from 2,000 to 8,000 of 10,000 sessions.
/diagnose-change Separate mix from within-channel change.
Don’t roll back signup yet.
Check the traffic first.
- −360 Signups lost to the shift: direct fell 6,000 sessions, paid rose 6,000.
- +100 Signups won back by higher rates in both channels. Real, but not this week’s priority.
- → 280 short of a 900 plan. First rule out a tracking relabel, then find where direct went.
/make-the-call Size the drop against the plan and recommend the next move.
Synthetic four-row CSV on GitHub · the 900 plan is a stated teaching assumption · panels show the target diagnosis and recommendation, not a model transcript
diagnose-change and make-the-call run the four-row case. Drop the --skill flag to install all fourteen.
npx skills add patrickjmorris/aftergrid --skill diagnose-change make-the-callPaste this into Claude Code or Codex. Codex’s workspace sandbox blocked the write when tested; here is what was checked.
Install the aftergrid analytics skills into this project.A number is the input.
The call is the output.
Pulling a metric is one step. A good analyst does the work before and after the query.
Frame the decision
“Did onboarding work?” is four questions. Find the one the owner is deciding.
grill-question02Explain the move
Reproduce it. Separate a traffic shift from a real change.
diagnose-change03Make the call
Size the move against the goal, weigh the options, and name what would flip the recommendation.
make-the-call04Pressure-test it
60% versus 40% is real. It is not a mandate.
analysis-reviewThen use the same skills on your own CSV, notebook, or warehouse.
What the agent refusesChoose a workflowSeven cases,
data and code included.
A conversion drop that was a traffic shift. MRR growth that hid $80 of churn. A causal headline on self-selected retention. A follow-up that re-tests last week’s lesson. A YouTube rule change sized on real creator data, and how it was built. Two real analyses of New York open data.
Explore the examplesConversion fell. Both channels improved.
The rollback meeting. Total down, both channels up.
Follow the analysisSynthetic lab · revenueMRR grew. The starting customers shrank.
Headline +$10. Starting customers −$80. Reconcile before you celebrate.
Follow the analysisSynthetic lab · lesson reuseConversion recovered. Was it the same mechanism?
Last time it was mix. Check before repeating the story.
Follow the analysisSynthetic lab · reviewRetention was 50% higher. Did onboarding cause it?
A mandate and a forecast sitting on a self-selected 60% vs 40%.
Follow the analysisReal data · Sports YouTubeSports YouTube’s fall 2026 will look like a record season. The new view count is the leading reason
Real data from 746 sports creator channels across 14 sports. Same creators, same video age: long-form views rose 74% after YouTube changed how it counts a view, in the same weeks the football and soccer seasons began. What it means for creators, clubs, leagues, and sponsors.
Agent-reviewed · not human-approvedFollow the analysisReal data · ProcessHow the sports YouTube deck was built
Six aftergrid skills took one real question from a raw view count to a 10-slide deck someone can act on. What each skill did, what it caught, and the move you can reuse on your own data.
Follow the analysisReal open data · NYCTwo questions about New York
Real taxi, weather, and bike data. One inconclusive answer, one descriptive one, dead ends kept.
Draft Findings · not publication-approvedFollow the analysisPick the skill
the work needs.
13 skills run on your existing files and SQL connections. setup-aftergrid adds the optional Engine. Not sure where to start? Run ask-aftergrid.
14 skills
ask-aftergrid
Choose the next skill for the work in front of you.
Framegrill-question
Turn a vague ask into a question worth answering.
Frameexplore-data
Learn the shape of the data before trusting it.
Framedefine-metric
Make the denominator, window and exclusions explicit.
Frameplan-analysis
Choose the comparison and the checks before the result.
Frameanalyze
Carry a question through to a reviewed answer.
Analyzediagnose-change
Reproduce a move, separate mix from within-group change, and test rival explanations.
Analyzewrite-finding
Turn the analysis into an answer someone can use, or put a buried answer first.
Explainiterate-visual
Make the chart earn the space it takes.
Explainmake-the-call
Tie a checked result to the goal, weigh the options, and recommend a call.
Decideanalysis-review
Try to disprove the answer before someone acts.
Improverevise-finding
Revise the answer without losing what supported it.
Improvelearn-from-analysis
Keep the lessons that should survive the analysis.
Improvesetup-aftergrid
Add the optional Engine for evidence and reproducibility.
Engine · optionalNo skills match. Try a different word or stage.
The habits
behind the skills.
Four essays on the judgment around the query.
The number is not the answer
Getting a metric is close to solved. Knowing whether it matters, and what to do about it, is the job.
Ask before you query
A clear question names the decision, the denominator, the comparison, and what would overturn the answer.
The metric moved. Now what?
Separate the arithmetic from the story.
Analysis that compounds
Keep lessons the next analysis actually retrieves.
Built by an analyst.
Available for hire.
I’m Patrick Morris. I ran data for ESPN+ as it grew to 25 million subscribers and built the experimentation platform at The New York Times. Agents got good at fetching the number. They still could not tell a product lead whether it mattered. I built these skills to close that gap.
About the projectAdvisory for data and analytics leaders.
Metric strategy, decision review, experimentation programs, and analytics agents your team can trust. For VPs and heads of data, analytics, and product.
Work with meSponsorshipFund the public skills and examples.
A named line here and in the docs. Nothing commercial inside a Finding, and no sponsor touches a metric approval.
Sponsor aftergridNext time a metric moves,
bring a recommendation, not just the number.
Install two skills and paste the four-row prompts. See how far your agent gets past the rollback.
Try the four-row case