Karthik Data Visualization Skills

Skills for Claude Code and Codex that make charts, tables and data stories the way I make them: one claim per chart, direct labels instead of legends, colour only where it means something, no chart furniture that isn't earning its keep, and a render-and-look check before anything is called done.

There is an optional local MCP server that does the arithmetic for colour assignment and rounding. The skills work without it.

Start here

Pick the entry point that matches what you have.

You have	Ask your agent to use	What happens
Data and a chart in mind	karthik-data-visualization	Builds the chart in the house style, renders it, and checks the export. Pulls in dataviz-selector if the form isn't settled.
An existing chart (image or code) that isn't working	dataviz-fix	Diagnoses it, rebuilds a real image, revises from your feedback until you accept it.
A dataset and a loose question	dataviz-orchestrator	Pins down the metric and denominator, cleans the data, finds the story, then builds and checks the chart.
A chart someone else made	dataviz-critique	Says what works, what misleads, and what to change.
A table to format	karthik-table-style	Alignment, rounding, emphasis and in-cell bars or shading.

For a single chart, karthik-data-visualization on its own is the whole thing.

Install in two commands (details in Quick start):

git clone https://github.com/skthewimp/karthik-data-visualization-skill.git
cd karthik-data-visualization-skill && ./sync.sh --no-pull --surface claude   # or: --surface codex

R with ggplot2 and ragg is the preferred renderer. Python and Matplotlib work as a fallback, but only with the typography, palette and spacing set deliberately; default Matplotlib output fails these skills.

How the pieces fit

dataset -> analysis-planner -> data-cleaning -> selector -> data-visualization (+ annotations) -> critique
chart image -> fix: critique + selector -> data-visualization -> inspect -> your feedback -> accept

dataviz-orchestrator owns the first sequence, dataviz-fix the second. Each loads only the skills the job needs.

Skill map

Start points

• karthik-data-visualization - chart craft: typography, direct labels, colour, axes, whitespace, export check.
• dataviz-fix - repair an existing chart through a feedback loop, then learn only reusable lessons.
• dataviz-orchestrator - dataset and loose question to finished visual story.
• dataviz-critique - chart critique and redesign pressure.

Choosing and checking

• dataviz-selector - which form fits the claim, including when a table beats a chart.
• dataviz-eval - send/revise verdict on a rendered chart, with every problem ranked. Only when you ask for one.

Craft details

• dataviz-color - choose and assign colours for this chart; brand first, then accessibility.
• dataviz-precision - how many digits to show, set by the spread of the numbers.
• chart-annotations - choose, rank, word and place annotations without clutter.
• chart-explainer - the two lines that travel with a chart in an email or notebook.
• karthik-table-style - tables as visualizations.
• karthik-powerpoint-style - claim-first, sparse analytical slides.

Before the chart

• dataset-question-generator - good questions from a raw dataset.
• karthik-analysis-planner - turn a fuzzy question into metric, denominator, comparison and falsifiers.
• karthik-data-cleaning - inspect, clean in context, inspect again.

R style

• karthik-r-analysis-style - how an exploratory R notebook is written.
• karthik-r-code-style - how R code is laid out.

Repository layout

.
├── <skill>/{codex,claude}/SKILL.md   # 17 skills, one folder each, a SKILL.md per client
├── dataviz_mcp/                      # Optional local stdio MCP: colour, precision
├── docs/                             # Human docs, one page per skill
├── sync-skills.py                    # Install Codex or Claude skill surfaces
└── sync.sh                           # Pull + install wrapper

Every public folder has a README. No generated dist/ tree is committed.

Quick start

Clone the repository and choose the skill surface for your client:

git clone https://github.com/skthewimp/karthik-data-visualization-skill.git
cd karthik-data-visualization-skill

./sync.sh --no-pull --surface codex   # Codex
# or
./sync.sh --no-pull --surface claude  # Claude Code

Install the MCP package into your existing Python environment:

python3 -m pip install -e .
MCP_PYTHON="$(python3 -c 'import sys; print(sys.executable)')"

Register it with Codex:

codex mcp add karthik-dataviz -- "$MCP_PYTHON" -m dataviz_mcp
codex mcp get karthik-dataviz

Or register it with Claude Code:

claude mcp add-json --scope user karthik-dataviz \
  "{\"type\":\"stdio\",\"command\":\"$MCP_PYTHON\",\"args\":[\"-m\",\"dataviz_mcp\"]}"
claude mcp get karthik-dataviz

Start a new client session after installation so it loads both the skill text and MCP tools. No daemon is required; the client starts the stdio process when needed.

Skill installation details

./sync.sh

This pulls latest changes and installs every skill to ~/.codex/skills/<skill> and ~/.claude/skills/<skill>.

To install without pulling:

./sync.sh --no-pull

To validate metadata without installing:

./sync.sh --no-pull --validate-only

To install one surface only:

./sync.sh --no-pull --surface codex
./sync.sh --no-pull --surface claude

MCP tools

The MCP server is optional. The skills work without it; the tools turn two by-eye decisions into arithmetic. Analytical and visual judgement stays in the skills, and charts are rendered and checked by looking at the export.

tool	what it computes
recommend_colours	which colour goes to which series, from brand or source colours, the background, a focal series and colour meanings
recommend_precision	one uniform rounding place for a column, from its spread

See docs/mcp.md for the boundary between tools and skills, and dataviz_mcp/README.md for installation, client registration and tool parameters.

Development notes

• AGENTS.md contains the maintainer-only validation and publish rule. It does not apply to third-party clones or forks.
• Source skills live in <skill>/{codex,claude}/SKILL.md.
• sync-skills.py discovers every root-level directory containing both surface files.
• sync-skills.py --validate-only checks frontmatter without copying files.
• For a localized fix, run the affected test files, optionally narrowed with -k: python3 -m pytest -q dataviz_mcp/tests/test_precision.py. Prose-only edits need metadata validation, not tests. See AGENTS.md for check selection.
• python3 -m pytest -q runs the full MCP suite in about a second.
• Extend an existing test when it already builds the same scenario. Keep separate cases for distinct failure modes; avoid wording snapshots and tests that merely repeat implementation constants.
• No generated dist/ output is committed.
• Keep README files in public folders. They are navigation aids for newcomers and should be updated when layout changes.

Session notes and writeups

• CHANGELOG.md - release-style summary of public repo changes.
• DEVLOG.md - session notes with prompts and work done.
• docs/blog/building-the-dataviz-selector-skill.md

License

MIT.