Community이미지github.com

visualize

Data storytelling for write-ups, reports, dashboards and slides. Plans the visual story across many datasets (order, merge, demote, cut, audience placement, consistent colors) and decides and draws each visual for Markdown, HTML and slide images, with headline, emphasis and checked numbers. Use when a document or deck needs charts, when choosing how to show a dataset, or when converting visuals between Markdown and HTML.

visualize란 무엇인가요?

visualize is a Claude Code agent skill that data storytelling for write-ups, reports, dashboards and slides. Plans the visual story across many datasets (order, merge, demote, cut, audience placement, consistent colors) and decides and draws each visual for Markdown, HTML and slide images, with headline, emphasis and checked numbers. Use when a document or deck needs charts, when choosing how to show a dataset, or when converting visuals between Markdown and HTML.

지원 대상~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/ndany/plotpoint/tree/HEAD/plugins/plotpoint/skills/visualize

Installed? Explore more 이미지 skills: steipete/songsee, affaan-m/frontend-design, mattpocock/improve-codebase-architecture · View all 6 →

즐겨 사용하는 AI에게 물어보기

이 에이전트 스킬이 미리 로드된 새 채팅을 엽니다.

문서

plotpoint: visualize

Purpose: better storytelling with data. Every choice here (what to show, in what order, where, how it's drawn) serves the question the reader needs answered. The caller owns the analysis, the prose and the final layout; this skill advises on the visual story and draws visuals. It never invents an analytical conclusion.

Two operations. Work out which one the caller wants, then follow only that section. Run uv run <this-skill>/scripts/visualize.py ... when uv is on PATH (it installs the dependencies itself). Otherwise use python3 (macOS, Linux) or py (Windows); the script finds its own cache venv once it exists. If the script stops with a missing-dependency message (exit code 3), show that message to the user and stop: don't install packages globally, don't look for another environment, don't work around it.

OperationInputOutput
storyboardA storyboard request: questions, audiences, datasets, targets, optionally the draft write-upAn ordered list of visuals with their intents, recommendations, a color map
renderOne visual: a dataset plus an intent from storyboard, or just a questionFiles, a Markdown snippet or HTML fragment per target, check results

Specs and field meanings: references/spec.md. Validate anything you write: visualize.py validate <kind> <file>.

Operation: storyboard

  1. Validate the request: visualize.py validate storyboardRequest request.yaml.
  2. Profile every dataset: visualize.py profile request.yaml. Read the profiles, not the raw rows. They include ranges, outlier and dominance flags, the rows holding extremes, and samples (references/profiling.md). For many or large datasets, do this step and step 4 in a sub-agent if you can, and keep only the result.
  3. If the request has a draft, read it: it shows the narrative, the current visuals and what the caller already concluded.
  4. Plan the story, per references/storyboard.md:
    • Map each question to the dataset(s) that answer it.
    • Order: summary visuals first, the one answering the main question leading.
    • Merge datasets that tell one story; demote visuals with no single takeaway to tables; cut datasets that answer no question.
    • Audience and placement: summary for decision-makers, body or appendix for practitioner detail.
    • For each visual, decide the intent (step 2 of render below). If it renders from a merged or derived table, set data and data_note.
    • Decide the rendering per target: the best technique in each format, with compensations where a format can't carry part of the intent (references/targets.md).
    • Build one color map for the whole document.
  5. Write the result as YAML and validate it (validate storyboardResult). Then dry-run every visual: visualize.py check-plan request.yaml result.yaml runs each one's checks (and shows text outputs) without writing files. Fix failures and re-run; then return it. Tell the caller which headlines are drafts and which recommendations need their decision.

Operation: render

  1. Validate the request (validate renderRequest).
  2. Decide the intent if the request has none (mode C). Profile the data (visualize.py profile request.yaml works on a render request, inline data included; or profile --data FILE), then:
    • Visual type from the question first, then the data's shape (references/chart-chooser.md). A table, a single number or a sentence is a valid answer.
    • Headline: a sentence stating the takeaway, not the topic. Use the caller's takeaway if given (headline_status: confirmed). Otherwise state what the data shows and mark it draft. A reference table with no single takeaway gets a descriptive title.
    • Fields: which columns play which role; display names for columns; bins for a histogram (a maximum, default 20; names: {count: ...} titles its y axis).
    • Highlight: the series or categories the headline is about. Everything else turns context gray.
    • Callouts: mark the moment that matters (rule, point, note; how each is pinned: references/spec.md, Callouts). Every number in the headline must be visible on the visual, so add a callout for any figure the chart doesn't plot (a total, a median). Highlighted series already get a value label; don't repeat it.
    • Outliers: if the profile flags any in a plotted measure, set outlier_handling, per column when they need different treatment: callout, log_scale, exclude (with values and a reason; only when the caller confirms), or accept (with a reason). Never drop one silently. Skew alone needs nothing; tables are exempt.
  3. Render: add the intent to the request and run visualize.py render request.yaml. Targets, layouts and techniques: references/targets.md.
  4. Read the checks in the result. Fix every fail (headline_numbers, outliers, callouts, legibility, render) and re-render; report warns. Read what each headline number matched: a pass on the wrong row, period or statistic is still wrong, so confirm each match is the one your sentence means. Render flags (--dry-run, --quiet): references/spec.md, "Render flags".
  5. Look at the output before delivering: open the PNG or SVG, or screenshot the HTML fragment (an SVG converts to PNG with vl_convert.svg_to_png, or opens in a browser); for text, tables and Mermaid, read the snippet. The chart file has no headline or source line; the snippet adds them. Check for clipped or overlapping labels, that every number in the headline is visible, and that the claim reads at a glance. Checks can't see layout.
  6. Return the snippets and files. Files carry a visible DRAFT or CHECKS FAILED mark when they aren't ready (references/spec.md, "Status marks"). Flag draft headlines and any compensation the caller should know about.

Rules that always apply

  • Rendering style follows the dataviz skill. The renderer's theme (config/theme.yaml) already applies its method, so render-only work needs nothing more. Load dataviz when you draw anything by hand (a type the renderer refuses, a dashboard layout); if it's unavailable, use references/fallback-rendering.md.
  • One y-axis. Bars start at zero. Text is ink-colored, never series-colored.
  • Source and as-of date on every visual (the renderer adds them).
  • HTML output is a fragment (<figure>), never a full page.
  • For multi-chart pages, also use references/composition.md.

관련 스킬