| name | dataviz-orchestrator |
| description | Orchestrate dataset-to-visual-story work: plan analysis, run it, choose visuals, style, critique, and iterate. |
Dataviz Orchestrator
Own the full loop from loose question to usable visual story:
dataset + loose question + audience
→ analysis contract
→ data profiling and analysis
→ facts table
→ story candidates
→ visual choice
→ chart implementation
→ rendered-output critique
→ iteration
→ final chart + notes
This skill sits above the existing analysis and dataviz skills. Load/use them at the right stage instead of duplicating their rules.
Required companion skills
Use these skills during the workflow when available:
karthik-analysis-planner: turn the loose question into measurable definitions, denominator, grain, comparison, caveats, falsifiers.r-analysis-rules: for R/tidyverse analysis code, notebooks, and pipelines.dataviz-selector: choose chart form, encodings, context layers, and what to avoid.karthik-data-visualization: apply Karthik's visual style and render-inspect-adjust loop.dataviz-critique: critique the rendered chart, not just the code, and propose fixes.babbage-visual-style: only when making Babbage-branded visuals.karthik-powerpoint-style: only when the final artifact is a slide/deck.
Operating defaults
- Do not ask for clarification unless the dataset/question/audience is impossible to infer. Make reasonable assumptions and state them.
- Prefer doing the analysis over writing a plan. The plan is a guardrail, not the output.
- Never let prose outrun computed evidence. Make a compact facts table before writing the claim.
- If the data cannot answer the question, say so and pivot to the closest defensible question.
- Prefer one strong interpreted visual over a dashboard.
- Keep outputs reproducible: leave analysis/chart code in a sensible project file unless the user asks for a one-off.
- Render charts and inspect the actual output image before calling the work done.
- Make the narrowest repo change needed. Do not regenerate unrelated artifacts.
Workflow
1. Intake and repo orientation
Identify:
- dataset path(s), format, source, and size;
- user's loose question or suspected pattern;
- intended audience and medium: notebook, blog, slide, social post, internal memo, or exploratory scratch;
- existing repo conventions: language, filenames, output folders, chart style, themes, and prior examples.
If the user gives no audience, assume: “Karthik exploring first, then possibly public-facing if the story is strong.”
2. Create a lightweight analysis contract
Use karthik-analysis-planner, but keep it compact unless the user asks for a full contract.
Must decide before coding:
- unit of analysis;
- denominator and numerator;
- key metric(s);
- primary comparison/baseline;
- filters/exclusions;
- sanity checks;
- what would weaken or falsify the likely story.
3. Profile the data
Run quick data checks before analysis:
- schema, row count, key columns, types;
- date/time coverage and timezone where relevant;
- missingness and duplicates;
- grain check: one row per what?;
- category cardinality and small-n groups;
- outliers/extreme values;
- whether the dataset contains the fields needed for the intended question.
Save or print only the useful profile summary. Avoid noisy dumps.
4. Analyse before charting
Build the smallest analysis table that can answer the question. Compute:
- main metric at chosen grain;
- denominator/sample-size columns;
- relevant baselines or comparison groups;
- sensitivity checks when definitions are fuzzy;
- a compact facts table with the numbers that would appear in title/annotations.
For exploratory work, try 2-4 plausible cuts, then choose. Do not produce a gallery unless the user explicitly asks.
5. Find the story
From the facts, write 2-3 candidate claims:
- Main claim: strongest defensible point.
- Alternative lens: a different useful comparison if the first claim is weak.
- Caveated claim: what can be said safely if data quality or causality is limited.
Pick one main claim for the first chart. If no claim survives, show the diagnostic result instead.
6. Select the visual
Use dataviz-selector.
Specify:
- recommended chart form;
- x/y/colour/facet/label encodings;
- ordering and scale choices;
- context layers: benchmark, event line, counterfactual, uncertainty, knee-bend, local maximum/minimum, annotation;
- what not to use and why.
Default to simple static visuals: line, sorted bars, scatter, small multiples, distribution, compact table/sparkline, or map only when spatial pattern matters.
7. Implement in Karthik's style
Use karthik-data-visualization and repo conventions.
Defaults:
- chart-first, low decoration, high data-ink;
- claim-first title;
- direct labels over legends;
- grey for context, colour for story emphasis;
- source/timeframe/definition note where needed;
- no decorative palettes, 3D, pies, donuts, gauges, radar charts, or dashboard clutter;
- stable descriptive output filename.
If using R, follow r-analysis-rules: tidyverse-first, %>%, concise notebooks/scripts, no noisy status chatter.
8. Render, inspect, critique, iterate
Render the chart to PNG/SVG/PDF as appropriate. Inspect the actual exported image.
Use dataviz-critique on the rendered output:
- Does the question-data-visual triangle hold?
- Is the main claim visible in 3 seconds?
- Are denominators, units, source, and timeframe clear?
- Are labels legible and non-overlapping?
- Does every colour/annotation earn its place?
- Is any conclusion overstated?
Iterate until fatal and major issues are fixed. Usually do 1-3 passes; stop when remaining issues are minor or require a different dataset/question.
9. Final response
Keep the final response short. Include:
- what artifact(s) changed/created;
- the final defensible story in one sentence;
- key caveat(s);
- path to output chart and code/notebook;
- any next chart worth doing, only if useful.
Output package
For a completed run, aim to leave behind:
- analysis/chart code or notebook;
- exported chart file;
- compact facts table, either in code output, CSV, or notebook cell;
- short note with the claim, definitions, caveats, and source.
Do less if the user asked for quick exploration; do more only if they asked for publication/deck-ready output.
Failure handling
- Dataset missing or unreadable: report the exact path/error and the next command/user action needed.
- Question not answerable: show which required field/comparison is missing and propose a nearby answerable question.
- Weak/no story: provide the negative finding or diagnostic visual rather than forcing a story.
- Crowded visual: split into small multiples, rank to top-N with context, or switch to table/sparkline.
- Causality gap: rename “effect” to “association/difference” unless there is a defensible design.
