The standard advice for getting more out of AI is still “write a better prompt.” That advice runs out fast. Prompting is the equivalent of hacking a fix straight into core: it holds for an afternoon, then you own it forever. If you want more than one-off code generation, Agent Skills in Data Science are the part worth learning.
From one-off prompts to modular skills
After 14 years of WordPress work and some genuinely ugly data migrations, I can say the bottleneck is almost never the code. It’s the repetitive logic that lives in my head and nowhere else. A skill is where that logic goes: a reusable package of instructions, metadata and supporting files that tells the model how to handle one recurring workflow.
So instead of pasting 5,000 lines of context into Claude or Codex every time you want a dataset looked at, you write a SKILL.md file. The model reads the metadata first and only pulls in the long instructions once it decides the skill applies, so the main context stays clean. Same idea as wp_enqueue_script: you don’t load the kitchen sink on every page, you load what the page needs.
The storytelling-viz skill that replaced a manual routine
A colleague, Yu Dong, had been publishing a weekly visualization for eight years, all of it by hand: find the data, open Tableau, build the chart, publish. That routine is now a storytelling-viz skill, and the model does the first pass, pulling out what looks interesting and proposing chart types against a fixed set of design principles.
I lost three days to the same lesson. I was trying to automate a WooCommerce sales report with raw prompts and the model kept hallucinating date formats, maybe half the runs. Moving that logic into a structured skill fixed it. The playbook is fixed now: check the up-funnel metrics, segment by region, and only then draw anything.
If you end up running several of these at once, it’s worth reading how to scale your parallel AI agent workflows using Git worktrees. That’s what keeps multiple skill iterations from stepping on each other.
How to structure your first agent skill
There isn’t much to it mechanically. The difference is that you’re writing down how the job gets done, not only what you want out of it. A skill directory usually looks like this:
/my-data-viz-skill
├── SKILL.md <-- Metadata and high-level instructions
├── templates/ <-- Reusable visualization templates
├── scripts/ <-- Data processing scripts (Python/JS)
└── examples/ <-- "Good" vs "Bad" output examplesThe SKILL.md itself needs a name, a description and the exact triggers that should fire it. If you wire in the Model Context Protocol (MCP), the skill can reach a database like BigQuery or your local SQL files on its own, which removes the copy-paste step in the middle.
Why prompting stops paying off
Raw prompts throw away everything you worked out last time. A skill is where domain knowledge gets written down. If you know that a slow site means checking wp_options for runaway transients first, that isn’t a prompt you should retype every month. Make it a “Performance Audit” skill and it runs the same way each time.
It also keeps the pieces separable. In the visualization example the process was split into two skills, one that generates and one that publishes. That’s the part I’d take away from mastering the AI data science workflow: you’re building tools you can reuse, not asking better questions.
If setting this up is eating hours you don’t have, I can take it off your plate. I’ve been dealing with WordPress and AI integrations since the 4.x days, so I know where the race conditions and the slow spots tend to hide.
What you get back from automating this
Automating 80% of a workflow doesn’t give you less work, it gives you different work. The tedious part stops being yours, and the time goes into the data itself and what it actually says. Whether you start from the public skills collection or write your own, you want a workflow that behaves tomorrow the way it behaved today.
