Mindmap Render

When to use this skill

Use this skill when the user asks to:

• Create a mindmap from a topic or data.
• Convert a Markdown outline into a visual mindmap image or PDF.
• Generate a colorful, presentation-quality mindmap with auto-export to PNG/PDF.
• Build a structured outline (unordered list) and then render it as a mindmap.

Prerequisites

Assume the runtime environment already has Python 3.10+, Playwright, and Chromium installed (they are provisioned in the container). Do not proactively run pip install or playwright install — just run the render script directly. Only if the first run fails with a clear missing-dependency error (ImportError, missing-browser error, etc.), then repair the environment:

pip install -r scripts/requirements.txt
playwright install chromium

and retry. Never install speculatively before a failure is observed.

Workflow

Step 1 — Determine input source

Ask the user (or infer from context):

• Topic: What is the central theme?
• Data source: Do they already have a Markdown file, or should you research and write one?
• Source fidelity vs synthesis — a spectrum, not a binary. The more specifically the request points to a named existing artifact (a particular book's table of contents (ToC), a particular course's syllabus, a specific spec or documentation structure, a numbered chapter list), the more you should reproduce the source's real structure verbatim — preserve original labels and numbering, follow the source's natural depth, and add NO fabricated descriptions. The more the request is a broad topic with no single canonical source, the more the structural targets below apply. Most requests sit somewhere on this spectrum; judge and lean accordingly.
• Theme: air (light blue glow + white cards, default), editorial (warm paper + jewel-tone branches), midnight (deep black + neon accents), or zen (soft misty background + muted pastels).
• Language:
• If the user explicitly specifies a language (e.g. "in English", "in Japanese"), use that language for all node text (labels + layer-5 descriptions).
• Otherwise, match the language of the user's request; default to English when the request language is unclear.
• Do not translate proper nouns, model names, or established technical terms — preserve them inline.

Default structural targets — for synthesis work only. These do NOT apply when you are faithfully reproducing a named source (in that case, follow the source's actual shape). They also yield to any numbers the user gives explicitly.

• 1 root + 6–10 top-level branches (default aim: ~9).
• Maximum depth = 5 layers (root → branch → subtopic → item → leaf-with-description). Layer 5 is reserved for the important, information-dense nodes — it is not a mandatory floor for every path.
• Layer-5 leaves (when present) MUST carry a substantive description — around 200 Chinese characters (or ~150 English words if the mindmap is in English), roughly 2–4 sentences — that explains mechanism, why it matters, quantitative detail, or a concrete example. A one-line label is not enough at layer 5; if you cannot write ~200 Chinese characters of real content, the node does not belong at layer 5.
• Intermediate nodes (layers 2–4) stay concise (1–10 words) and may themselves be terminal leaves when that is the right level of detail.
• Asymmetry is required, not a flaw. Branches should be weighted by importance and information value, not padded for visual symmetry:
• Pillar branches (where the real substance lives) should go deep and wide, with many children and rich layer-5 descriptions.
• Supporting / well-known branches can stop at layer 2 or 3. Do not expand common knowledge the target reader already owns, and do not invent filler children just to match sibling counts.
• When deciding "expand or stop," ask: Would a knowledgeable reader learn something here? If no, prune.
• The hierarchy is intentionally irregular. In Step 2c, report the shape honestly rather than forcing every branch to layer 5.

If the user gives different numbers, use theirs; otherwise treat the defaults above as guidance with judgment — breadth/depth targets are firm, but per-branch expansion is deliberately uneven.

Step 2a — User provided a Markdown file

If the user already has a .md file, note its path and proceed to Step 3.

Step 2b — Generate the Markdown outline yourself

CRITICAL: If the user has NOT provided a .md file, you MUST perform web research BEFORE writing the outline. Do not rely solely on internal knowledge.

1. Research (mandatory): Use WebSearch to find authoritative, high-quality sources:

• Official book table of contents (publisher's catalog, Douban Books listing).
• Wikipedia structured sections.
• Academic course syllabi or reputable blog series.
• Official documentation / white-paper outlines.
• Recent industry reports, survey papers, or conference proceedings (e.g. NeurIPS, ICML, JPMorgan Quantitative Research).

Fail loudly if the authoritative source cannot be found. When the user names a specific artifact (a particular book, edition, course, spec) and repeated searches do not surface its real ToC / syllabus / structure, STOP and tell the user: "I couldn't find the authoritative structure of [X]. Please paste the ToC, confirm the edition/title, or allow me to produce a synthesized overview instead." Never invent chapter/section structure to fill the gap. This is the single most important rule of this step.

1. Write the outline — choose the mode based on Step 1's fidelity-vs-synthesis judgment:

a) Faithful reproduction (user pointed to a specific named artifact and you located its real structure): copy the ToC/outline verbatim into a single-root bullet list. Preserve original labels, chapter numbering, and natural depth. Do not add layer-5 descriptions, do not force 6–10 top-level branches, do not pad to 5 layers — follow whatever shape the source actually has. The only transformations allowed are: wrapping everything under one root node, and cleaning trivial typography (e.g. converting full-width numbers consistently).

b) Synthesis (broad topic, multi-source): distill the research into a single-root unordered-list Markdown file following the Step-1 structural targets.

• Use standard - bullet lists; nesting = depth.
• One top-level bullet = the root (layer 1).
• Layers 2–4 (branch / subtopic / item): concise labels, 1–10 words each. A layer-2/3/4 node can be a terminal leaf when no deeper breakdown adds value.
• Layer 5 (when used): a substantive description, ~200 Chinese characters (or ~150 English words for English mindmaps), covering mechanism + why it matters + concrete detail (numbers, names, example). Only create a layer-5 node when you have real content of that density; never pad.
• Weight by importance. Give the pillar branches many children and deep layer-5 content; let well-known or low-information branches stay shallow. Target reader: an informed practitioner — skip what they already know, dwell on what is surprising, recent, or load-bearing.
• Irregular depth is expected. A tree with 3 deep pillar branches and 6 shallow supporting ones is healthier than 9 uniformly-expanded branches full of filler.

1. Save the file:

• Save to mindmap-output/.md (or the current project directory).
• Show the user the saved path and the first ~30 lines of the outline for confirmation.

Step 2c — Self-check before rendering

If you wrote a faithful reproduction (Step 2b-a), skip the full audit. Just verify two things and report one line each: (1) the outline's labels and numbering match the source, (2) you did not inject any fabricated descriptions or extra layers. Then proceed to Step 3.

If you wrote a synthesis (Step 2b-b), do not skip. After saving the outline and before running the render script, verify against the structural target. Output a short audit block to the user:

Structure audit:
- Top-level branches:    (target 6–10)
- Max depth reached:     (ceiling 5)
- Pillar branches (reach layer 5 with substantive content): 
- Shallow branches (stop at layer 2–3 by design): 
- Layer-5 leaves with ≥~200 Chinese characters description:  / 
- Shape note: 

Red flags — rework the outline before rendering if any apply:

• Layer-5 leaves that are one-line labels or under ~100 Chinese characters → either enrich them to ~200 Chinese characters of real content, or demote the node to layer 4.
• Every branch reaches the same depth with similar child counts → you are padding for symmetry; prune the weakest branches back.
• A branch exists only to list common knowledge the target reader already owns → cut it or collapse it.
• Pillar branches are shallower than supporting branches → rebalance so information density follows importance.

Step 3 — Render the mindmap

Run the rendering script:

python scripts/generate_mindmap.py \
  --md  \
  --output-dir ./mindmap-output \
  --title "" \
  --theme  \
  --scale 2

Arguments:

• --md (required): Path to the Markdown file.
• --output-dir: Where to place the results. Default is ./mindmap-output.
• --title: Used for the HTML `` and the output base file name.
• --theme: air (designer-style airy blue glow + white rounded cards + soft pastel branch accents), editorial (magazine-style warm paper background + jewel-tone branches + dark serif text), midnight (pitch-dark background + neon accents + crisp light text), or zen (soft misty background + muted Morandi pastels + gentle serif text).
• --scale: Upscale factor for the exported image (default 2). 1 = compact (~2–3 MB), 2 = crisp readable (~3–5 MB), 3+ = poster size. Larger numbers produce physically larger, more readable text.

Example (default shape from Step 1: ~9 branches, 5 layers, leaf descriptions):

python scripts/generate_mindmap.py \
  --md mindmap-output/large-test.md \
  --output-dir ./mindmap-output \
  --title "Artificial Intelligence Panorama" \
  --theme air \
  --scale 3

Step 4 — Deliver results

Report the three generated files to the user:

1. {title}.html — interactive mindmap (open in browser to zoom/pan/collapse). Live rendering: after starting an HTTP server in the same directory (e.g. python -m http.server) and accessing it through a browser, edit the .md file and refresh the page to see the update; opening it directly as a local file uses the embedded content, behaving the same as before.
2. {title}.png — high-resolution full-page image (suitable for slides, social media, docs).
3. {title}.pdf — vector-like PDF export with print background.

Important notes

• Color system — "rainbow branches" with in-family shading. Each top-level branch owns one color family (hue); its descendants use the same hue with depth-based variation (deeper layers → slightly lighter + less saturated on light themes; slightly dimmer + less saturated on dark themes). You do not need to configure this — it is applied automatically by the render script based on the theme palette.
• Do not pass raw paragraphs as the Markdown input. The renderer works best with bullet-list outlines. If the source text is prose, convert it into a hierarchical bullet list first.
• The script automatically strips YAML frontmatter from the Markdown file so markmap can focus on the outline.
• If the mindmap is very large, Playwright will resize the viewport to fit the entire diagram; full_page=True guarantees the PNG captures everything without clipping.
• Never skip research when the user only gives a topic. The mindmap's quality depends on accurate, up-to-date, well-sourced hierarchies.
• When researching, prefer sources that already have a clear hierarchy (ToCs, syllabi, wiki sections) so the resulting mindmap is accurate and useful.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ai4s-research
• Source: ai4s-research/ai4s-skills
• License: MIT

Install and usage instructions live in the source repository linked above.