Set up before you start
Put an environment phase at the front of the skill. "If you need an environment set up, you can just tell it, 'Run these commands.'" No mid-run dependency crashes.
Show Us Your Agent Skills / EP 01 / guest dossier
RANDY OLSON
GOODEYE LABS
R/DATAISBEAUTIFUL
TUFTE TEST
TOUCHING GRASS
RANDY OLSON
Randy is teaching an agent his data visualization judgment: how to research, what makes a chart good, and how to check its own work with an LLM judge applying Tufte's principles to the image. He runs the skill every morning and makes it better after every run.
1/DAYa chart every morning for his post series
10MINfor research that once took him a whole weekend
3+chart variants before committing to a direction
TUFTEprinciples encoded as an llm judge on the image
4tabs running while he goes for a walk
+1/RUNthe skill learns something from every run. it compounds
this episode is also a blog post: The Agentic Software Factory and a podcast: Agentic Engineering and the Lost Art of Verification
EP 01 · RANDY OLSON · the chart workflow, live on stream
TASTE, ENCODED
the high-signal-chart-workflow. install it, then make it yours
"What popped out was honestly unimpressive. And then that's when I realized, oh, I have to encode still what makes for a good data visualization, at least in my opinion." The skill carries the taste Randy built over years of making data topics accessible: no chart junk, a clear story, clear annotations, straightforward color, and research biased toward CDC, government, and educational sources.
One line in, one chart out: the agent researches the web, turns scattered sources into a dataset, tries several chart types, checks the result with code and an LLM judge, and opens chart.png for Randy's review. The taste that can't be encoded yet, like catching "this freaking overlap," stays his job.
Inside the skill: the full SKILL.md, how to run it in Claude Code, and how to swap in your own taste.
"You don't want to just tell it what to do, you also want to tell it how to check it."
The verifier loop is guardrails on open-ended generation: code checks for things like DPI, an LLM judge applying Tuftean principles to the image, pass or fail. 01:29:30
ONE LINE TO CHART.PNG
the workflow, start to finish. every timestamp opens the segment
Start from one line"Visualize the history of marriage and divorce in the USA," typed into Claude Code with the workflow template pulled in. 01:26:39
Set up the environment firstA setup phase verifies libraries before any code is written, so the agent doesn't crash into missing dependencies mid-run. 01:15:25
Research toward reliable sources"It took an idea, researched the web, and then turned it into a dataset." Data gathering that once took him a weekend of PDF scraping. 01:28:01
Generate variants before committingLine charts, small multiples, area charts: the first act of visualization is understanding the data's useful forms. 01:28:29
Verify with code and a judgeThe Tufte test looks directly at the image and says pass or fail. If it fails, the judge gives feedback and the agent keeps fixing. 01:29:44
Randy closes the loopHe opens chart.png, picks his variant over the agent's, catches "this freaking overlap," and feeds the lesson back into the skill. 01:33:00
"They're too agreeable by default, and you really have to fight them to get them to not be agreeable."
His digital twin context tells agents he wants pushback: "someone who's not necessarily harsh, but won't just agree with me." 01:10:28
A SKILL IS A PROGRAM
his four rules for skills that repeat, from the live walkthrough
Design a thin driver
Long skills decompose into per-phase reference files behind progressive disclosure. A run that starts at phase four doesn't carry phases one through three in context.
Be exact where it matters
Exact steps, exact commands, exact code snippets: "the more specific you can be in your skill, the more repeatable it is in the future."
Reflect and improve every run
The final phase asks how the run went and what to fix. "Every single run, you're learning something new and putting it into the skill. It's compounding."
"I would love it if they trained on my skill and then we didn't need the skill anymore."
Models and harnesses are mostly fixed. "Skills are really the thing that can evolve with you," and their traces can become the next model's training data. 01:22:42
THE STACK
the chart series runs on claude code. the taste comes from tufte