| name | viz |
| description | Tufte-style visualization critique/design skill. Use when user says "/viz", asks to design or improve charts, dashboards, data tables, comparison views, scorecards, cockpit UIs, analytical HTML artifacts, trading/analytics dashboards, idea-ranking cockpits, feed-curation tasklists, or review-queue views. Applies data-ink ratio, chartjunk removal, graphical integrity, small multiples, sparklines, layering, provenance, and high-density/low-slop defaults. |
| allowed-tools | Read, Edit, Write, Bash |
Viz — Tufte-Style Analytical Visualization
Overview
Use this skill to design or critique visual displays where the point is to understand evidence, compare alternatives, reveal signal, or make a cockpit/dashboard easier to reason with. It adapts Edward Tufte's visualization principles — graphical integrity, data-ink ratio, chartjunk removal, small multiples, sparklines, layering/separation, and high-density analytical design — to Second Brain artifacts and vibe-coded tools.
The stance: pretty is subordinate to truthful, dense, useful, comparison-rich. A chart or dashboard should help the owner decide what is good, what changed, what matters, and what to do next.
When to Invoke
Invoke when:
- User says
/viz.
- User asks to design, improve, critique, or clean up a visualization.
- User asks for a dashboard/cockpit/table/scorecard/HTML artifact that carries data or comparisons.
- User asks why a chart feels misleading, noisy, thin, hard to scan, or “startup-dashboard bullshit.”
Use for:
- Trading / analytics dashboards: alpha/returns, replay comparisons, source weights, sectors, model-classification changes, portfolio/risk views.
- Signal / bookmark / feed dashboards: tasklists, trend scans, source quality, recurring topic clusters.
- Idea-ranking / curation cockpits: taste-based comparison views, ranked candidates, review queues, decision surfaces.
- Story or content review tools: A/B comparisons, human-review queues, scorecards, benchmark views.
- HTML artifacts, reports, charts, data tables, diagrams, sparklines, or any visual analytical display.
- Critiquing existing visuals for misleading scales, decorative slop, insufficient comparison, or poor information density.
Do not use for pure brand/aesthetic pages with no analytical content; use design/style skills instead.
Core Principles
- Show comparisons. Every display should answer: compared to what?
- Show causality, mechanism, or structure where possible. Move beyond “number went up.” Show why, drivers, dependencies, or uncertainty.
- Respect graphical integrity. No fake magnitude. Baselines, intervals, area/volume encodings, and axes must match the data.
- Maximize data-ink within reason. Remove ornament, heavy grids, redundant labels, and startup-dashboard decoration that does not carry information.
- Prefer dense, readable displays over sparse bullshit. High information density is good when hierarchy/layering is clear.
- Integrate words, numbers, and images. Put labels and explanations near the data they explain.
- Expose provenance. Analytical displays should state data source, time window, filters, and caveats.
- Preserve decision context. Show enough surrounding data to prevent overreacting to a single metric.
Workflow for New Visualizations
1. Clarify the data story
Ask or infer:
- What decision does this support?
- What comparisons matter?
- Who is the viewer?
- What is the time window?
- What is the failure mode if the display misleads?
2. Choose the display form
- Time-series → line chart, sparkline stack, or annotated timeline.
- Many comparable entities → sorted table with sparklines / compact bars.
- Before/after or A/B → paired columns, slopegraph, small multiples, or delta table.
- Distribution → histogram, beeswarm, box/violin only if the audience benefits.
- Part-to-whole → usually avoid pie charts; prefer sorted bars or table with percentages.
- Multivariate signals → small multiples, faceting, layered table, or compact scorecard.
- Ranked review queue → sortable table with the few highest-signal fields, plus reason/evidence column.
3. Design with data-ink discipline
- Start minimal.
- Add only elements that carry data, comparison, hierarchy, provenance, or actionability.
- Use color sparingly for semantic states: positive/negative, selected/highlighted, stale/warning.
- Prefer subdued grids, direct labels, and compact legends.
- Default to tables when exact values and comparison both matter.
4. Add context and caveats
Always include when relevant:
- Source
- Date/time window
- Filtering rules
- Missing data caveats
- Whether values are raw, normalized, model-generated, or manually reviewed
Workflow for Critique
-
Graphical integrity
- Are scales truthful?
- Are baselines appropriate?
- Are intervals consistent?
- Is area/volume encoding lying?
- If proportions look suspicious, calculate/estimate lie factor.
-
Comparison quality
- Does it answer “compared to what?”
- Are deltas visible?
- Are cohorts/time windows comparable?
- Is ranking meaningful?
-
Data-ink / chartjunk
- What can be erased without losing information?
- Are decorative gradients, shadows, cards, icons, or 3D effects obscuring the signal?
- Are labels redundant or too far from the data?
-
Density and hierarchy
- Is the display too sparse for the question?
- Is it dense but illegible?
- Can small multiples, sparklines, or grouped tables improve scanability?
-
Actionability
- Does it make the next action obvious?
- If it is a cockpit, does the top of the page show what changed, what matters, and what needs review?
Owner-Specific Defaults
- Analytics dashboards: Prefer deterministic comparisons and replay deltas over vibes. Show before/after prompt/config changes, alpha/returns, source/sector effects, and relevant exclusions.
- Second Brain / idea-ranking cockpits: Preserve exact titles/keywords; show why an item surfaced. Provenance beats polish.
- Messaging / mobile output: Many messaging platforms have no real table syntax. Use bullets, labeled key/value rows, or compact lists. Do not emit pipe tables unless the target is a markdown file.
- HTML artifacts: Dense dark-mode cockpits are fine, but visual hierarchy must be earned. Use subdued chrome; make the data primary.
- Review queues: Put “why this matters” and “next action” near the item, not hidden below the fold.
Quick Checklist
Before finalizing a visualization or critique, verify:
References
references/tufte-principles.md — core principles from The Visual Display of Quantitative Information: lie factor, data-ink, chartjunk, small multiples, integrity.
references/analytical-design.md — material from Envisioning Information, Visual Explanations, and Beautiful Evidence: analytical design, sparklines, layering & separation, micro/macro, range-frames, causality, confections.
Common Pitfalls
- Making it pretty instead of useful. Aesthetic polish is secondary to evidence, comparison, and actionability.
- Under-dense dashboards. Four giant cards can be worse than one compact table when the owner needs to scan many candidates.
- No comparison baseline. A metric without history/cohort/benchmark often becomes decorative numerology.
- Hiding provenance. If the viewer cannot tell where the data came from and when it was captured, trust collapses.
- Messaging-platform tables. Pipe tables degrade in many messaging contexts. Use bullets/key-values unless writing to a file.
- Color abuse. Color should encode meaning, not serve as startup confetti.
- Over-reducing multivariate decisions to one score. Scores can rank, but include the reason/evidence fields that let the owner apply taste.