Skip to content

Releases: SNL-UCSB/data-visualization-skill

v2.0.0 — Six-Phase Workflow

27 Mar 16:43
@agupta13 agupta13
v2.0.0
e42dc6c

Choose a tag to compare

View all tags
v2.0.0 — Six-Phase Workflow Latest
Latest

data-visualization-skill v2.0.0

Major expansion from 3 to 6 phases, based on feedback from Walter Willinger. The core insight: students need to look at the data before forming hypotheses, and need substantive feedback on their figures, not just formatting help.

Part of a research skills family

This skill is one of three Claude Code skills built by the Systems and Networking Lab (SNL) at UC Santa Barbara, covering the core research pipeline:

  • literature-survey-skill — paper ingestion, claim extraction, landscape mapping
  • data-visualization-skill (this repo) — plot formatting, code generation, figure analysis
  • paper-writing-skill — section structure, voice consistency, page compression

All three share a common intellectual spine: compress the operational middle, protect the thinking (see Systems for Agents, Agents for Systems). The skills are independent but reinforce each other — survey findings feed paper positioning, visualization exploration reshapes defensible claims, and writing sharpens both reading and analysis.

Further reading:

What's new

Three new modes:

  • /viz ingest — CLI-first data inspection. Before opening Python, use Unix one-liners (head, wc, cut, sort | uniq -c) to get an unmediated sense of the raw data. Inspired by Philipp Winter's Research Power Tools. Saves a data_summary.md.

  • /viz explore — Disposable seaborn plots to build intuition before committing to a figure design. Concrete recipes organized by question: distributions (histplot, ecdfplot), relationships (pairplot, scatterplot, jointplot), group comparisons (boxplot, violinplot), time series (lineplot). A valid outcome is discovering the data is flawed and the experiment needs revisiting. Saves an exploration_log.md.

  • /viz analyze — First-principles feedback on the generated figure. The student interprets first (so the skill doesn't anchor their thinking), then the skill provides substantive analysis through three lenses:

    • Tukey: prediction vs. reality, residual thinking, re-expression
    • Tufte: data-ink compliance audit, chartjunk detection, scale integrity, over-summarization warnings, small multiples recommendations
    • Hypothesis validation: cross-reference against braindump predictions, skeptical reviewer stress-test

Other changes:

  • /viz execute is now purely operational (code generation + execution). All reflection moved to /viz analyze.
  • Experienced students can skip early modes and enter directly at /viz plan — the early modes are teaching tools, not gates.
  • New reference/cli_data_inspection.md — copy-paste Unix one-liners for data inspection.
  • Expanded Tufte section in README with specific principles mapped to skill defaults.
  • Credit added for Philipp Winter (Research Power Tools) and the seaborn tutorial.
  • WALTER attribution corrected (common practice in systems/networking research, not a lab-specific invention).

Workflow

Ingest → Explore → Brainstorm → Plan → Execute → Analyze
  ↑                                                  |
  └──── go back if data is flawed or hypothesis ←────┘
        needs revision

Artifact chain

Mode Output
Ingest data_summary.md
Explore exploration_log.md + disposable plots
Brainstorm braindump.md
Plan plot_context.md
Execute figure + Python script
Analyze WALTER narration + first-principles feedback appended to plot_context.md

Upgrading from v1.0.0

The v1.0.0 three-phase workflow (Brainstorm → Plan → Execute) is still available at tag v1.0.0 for those who prefer the simpler version. To upgrade, re-run ./setup.sh.

Installation

git clone https://github.com/SNL-UCSB/data-visualization-skill.git
cd data-visualization-skill
./setup.sh
Assets 2
Loading

v1.0.0 — Initial Release

27 Mar 00:40
@agupta13 agupta13
v1.0.0
e531309

Choose a tag to compare

View all tags
v1.0.0 — Initial Release

data-visualization-skill v1.0.0

A Claude Code skill for research-quality data visualization with a three-phase workflow: Brainstorm → Plan → Execute.

What's included

  • /viz brainstorm — Socratic questioning to clarify visualization intent before any code is written. Saves a structured braindump.md for durable context across sessions.
  • /viz plan — Translates intent into concrete design decisions (plot type, layout, data source, visual encoding). Produces a plot_context.md contract for execution.
  • /viz execute — Generates publication-quality Python code, runs it, and forces WALTER reflection (hypothesis → axes → look here → trend → exception → result).

Reference materials (usable independently)

  • reference/before_you_plot.md — 15-question pre-plotting questionnaire (Tukey-inspired)
  • reference/plot_context_template.md — Blank template for structured figure planning
  • reference/matplotlib_defaults.py — Drop-in module with SIGCOMM/NSDI-quality rcParams, golden ratio sizing, CDF/CCDF/time series helpers, and colorblind-safe palettes

Installation

git clone https://github.com/SNL-UCSB/data-visualization-skill.git
cd data-visualization-skill
./setup.sh

Then use /viz brainstorm, /viz plan, /viz execute in any Claude Code session.

Loading