Mindmap (markmap) Skill

Render hierarchical Markdown as an interactive SVG mind map with markmap.js, plus three custom layers: white-font CSS, expand-by-level control, and search that filters the tree.

source.md  ──►  filter / set expand level (Python)  ──►  build_html()  ──►  vendored markmap libs (local, offline)  ──►  SVG + toolbar
 (outline)        (manipulate the text)                  (HTML + CSS)        (transform + render in the browser)          (screen)

Helper functions live in scripts/render_markmap.py; a minimal source file is in assets/example.md; regression tests are in evals/. Deeper docs: references/internals.md (how the helpers work) and references/lessons.md (real-world lessons + the adversarial counter-review behind the current code).

When to use this skill

• The user asks for a mind map, markmap, or concept map of anything.
• Hierarchical content — an outline, notes, a plan, a document's structure, a taxonomy — should become navigable, scannable, or shareable.
• A map must be sent to someone or opened with no setup: the output is one self-contained .html that works offline.
• A mind map should be embedded in a Streamlit app.
• An existing markmap renders blank, white-on-white, or truncated — this skill carries the known fixes (§2).

---

1. Source format

markmap derives hierarchy from headings (#) and nested list items — -, *, +, or numbered (1. / 1)) — indented by 2 spaces per level. (Tabs work too; the filter treats one tab as one level.) A YAML frontmatter block controls behavior:

---
markmap:
  colorFreezeLevel: 2      # freeze color from level 2 down (branches keep the parent color)
  initialExpandLevel: 2    # levels kept open; root is level 1, so 2 = root + branches (-1 = all)
  maxWidth: 380            # max node width in px (forces wrapping)
---

# Root                     <- root (level 1)
## Branch A                <- level 2
### Sub-branch A1          <- level 3
- Leaf                     <- level 4 (bullet under a level-3 heading)
  - Detail                 <- level 5 (bullet indented +2 spaces)

Presets (apply_presets). You don't have to hand-write the markmap: block — apply_presets(src, color=None, max_width=380) fills sensible defaults without overriding anything you set: colorFreezeLevel: 2, maxWidth, and an initialExpandLevel sized by node count (expand-all for ≤30 nodes, else level 2). Pass color=["#7fd1ff", "#ffd479"] for a custom palette. Any key already in your frontmatter wins, so you can set one value and let presets fill the rest.

Content rule — term → parent / description → child. Put the label on the node and its explanation as a child, not on one line. Prefer:

- Term
  - description of the term

over - Term — description. This keeps nodes short and the tree scannable.

Node text may contain <, >, and & freely (a < b, List<String>, even </div>). build_html HTML-escapes the source before embedding it and the browser decodes it back, so markmap sees exactly what you wrote. (Earlier versions broke on a literal <; that footgun is gone — don't pre-escape to < yourself or it shows up literally.)

---

2. Rendering (white font)

The renderer is: the vendored markmap <script>s (local files), an <svg class="markmap"> plus a hidden source <div> holding the Markdown, a small init script (transform → Markmap.create → toolbar), and the CSS. See build_html() / render_markmap() in scripts/render_markmap.py, with the rationale in references/internals.md.

Non-obvious points (these cost rework):

• Offline, single self-contained file by default. build_html inlines the vendored markmap stack (d3 + markmap-view/-lib/-toolbar, pinned exact in assets/vendor/) directly into the HTML — no CDN, no network, and no sibling files — so the one .html opens offline anywhere you move or share it. (That single-file default is the fix for maps that failed to load their libs when opened alone.) Pass inline=False for a smaller HTML that instead references a vendor/ folder via vendor="vendor", which must then travel beside it.
• LaTeX math is not rendered in the offline bundle. markmap's KaTeX support needs window.katex, which isn't vendored, so $...$ / $$...$$ show as plain text. Everything else renders fully offline.
• Navigation toolbar (bottom-right): zoom in/out, fit-to-window, expand-all, collapse-all, and download as SVG / PNG. Pass toolbar=False to omit it. Expand/collapse set each node's fold then re-render with setData() and no argument — passing data re-derives fold from initialExpandLevel and would wipe the manual fold. Export snapshots the current fold state; the SVG inlines the white-font CSS + a dark backdrop so it stands alone, and PNG rasterizes at 2× (falling back to SVG if a browser refuses to rasterize the <foreignObject> labels — markmap draws node text as HTML-in-SVG, not <text>).
• White font is invisible without a dark background. This is the #1 way the map "renders blank": the font is white, the surface is white, so nothing shows. A standalone .html opens on the browser's white default, and a Streamlit components.html iframe is white by default too — neither inherits the host's dark theme. So build_html paints its own dark backdrop (background="#0e1117" by default). Only pass background="transparent" when you know the host behind the iframe is already dark and you want a seamless blend. White font + dark background travel together — never set one without the other.
• White font needs TWO selectors + !important. markmap draws text as SVG <text> and sometimes as <foreignObject> (HTML inside SVG). Style both or half the labels stay dark:

  svg.markmap text { fill: #ffffff !important; }
  svg.markmap foreignObject, svg.markmap foreignObject * { color: #ffffff !important; }
  

• Embed via an isolated iframe (components.html in Streamlit, or a standalone .html file). Host CSS won't leak in and the map's CSS won't leak out, so the <style> goes inline in the HTML string.
• The source is HTML-escaped before embedding. The init script reads the Markdown from the hidden source div's textContent, which the browser decodes — so escaping round-trips losslessly. Without it, a </div> or List<String> in the outline closes the div early and silently truncates the map.

---

3. Expand-by-level control

Don't rebuild the map — just rewrite initialExpandLevel in the frontmatter before rendering. Use set_expand_level(src, level) from the helper module:

• level 1/2/3 expands that many levels.
• level = -1 expands everything (the "expand all" button).
• Injection is scoped to the frontmatter and done in place: rewrite an existing initialExpandLevel, else add one under the markmap: key (block or inline form), else prepend a minimal frontmatter — but only when none exists. It never rewrites the word "initialExpandLevel" sitting in your body text, and never stacks a second --- block on top of existing frontmatter (markmap reads only the first block, so stacking would silently drop your other settings).

---

4. Search that filters the tree

When there is a query, keep only nodes that match + their path to the root (ancestors) + their subtree (descendants), so a hit appears in context instead of floating alone. Algorithm in filter_markmap():

1. Parse each line into a node. Heading level = number of #. List-item level = (last heading level) + 1 + indentation, where indentation counts 2-space or tab units (expandtabs), and the marker may be -/*/+/numbered.
2. Build an explicit parent tree with a kind-aware stack — a heading is parented to the nearest shallower heading, never to a bullet, even when its #-rank numerically exceeds a bullet's level. (Heading rank and bullet indentation share one number line, so comparing raw depths mis-nests an H4-after-a-bullet; the parent tree is what keeps ancestry correct.)
3. Mark each node match / no-match with an accent-insensitive comparison (_norm).
4. Keep every match, every ancestor (walk parent pointers up), and every descendant (any node whose ancestor chain hits a match). Rebuild the Markdown and force initialExpandLevel: -1.

_norm strips accents via Unicode NFD so "compliance", "COMPLIANCE", and accented variants all match. Zero matches yields a frontmatter-only (blank) map by design — callers branch on the returned count.

---

5. Minimal usage

The helpers live in this skill's scripts/ directory — and the working directory is normally the user's project, not this folder, so a bare "scripts" on sys.path won't resolve. Build paths from the skill's base directory (the path announced when this skill loaded):

import sys
from pathlib import Path

SKILL_DIR = Path(r"<this skill's base directory>")   # announced when the skill loaded
sys.path.insert(0, str(SKILL_DIR / "scripts"))
from render_markmap import write_mindmap, apply_presets, set_expand_level, filter_markmap

Single self-contained file (recommended) — writes the .md source of truth and a fully inlined .html that opens offline anywhere, with no sibling vendor/:

src = Path("outline.md").read_text(encoding="utf-8")  # the user's outline (sample: SKILL_DIR / "assets/example.md")
src = apply_presets(src)                       # fill default markmap options (no override)
# optional: src, n = filter_markmap(src, "branch b")   # search + keep context
# optional: src = set_expand_level(src, -1)             # expand all
write_mindmap(src, "out/mapa.html")            # -> out/mapa.md + a self-contained out/mapa.html
# smaller HTML + a sibling vendor/ folder instead: write_mindmap(src, "out/mapa.html", inline=False)

Just the HTML string (e.g. to embed) — build_html(src) returns one self-contained document with the libraries inlined:

from render_markmap import build_html
open("mindmap.html", "w", encoding="utf-8").write(build_html(src, height=850))

Inside Streamlit:

sys.path.insert(0, str(SKILL_DIR / "scripts"))       # SKILL_DIR as in the setup above
from render_markmap import render_markmap, set_expand_level
render_markmap(set_expand_level(src, level), height=850)

---

6. Authoring guidelines (balanced, scannable maps)

The shape of a mind map carries meaning, so structure the content deliberately:

• 5–8 main branches off the root. Fewer feels thin; more is hard to scan at once — group related ideas under an intermediate node instead of widening the root.
• 3–6 children per branch, kept roughly even across branches. One giant branch beside several stubs reads as unfinished; rebalance or split it.
• Labels of 1–3 words. A node is a handle, not a sentence. Put the explanation on a child node (the term → parent / description → child rule from §1), never inline on the label.
• Balanced depth. Keep branches within ~1 level of each other; a single branch that plunges several levels deeper than its siblings unbalances the layout.
• Let apply_presets set the frontmatter (color, wrap width, expand level) so you don't hand-tune each map (§1).

Quality checklist (run before shipping a map)

• Root has 5–8 branches; none is a dumping ground.
• Every branch has 3–6 children, roughly balanced.
• Labels are 1–3 words; long text lives on child nodes, not labels.
• Depth is even across branches (no lone deep tunnel).
• apply_presets applied, or the frontmatter set deliberately.
• Opened the .html with the network off (and from a different folder) — it still renders (self-contained).
• Searched a known term — it filters to that node + its context.

---

The 5 lessons this skill must carry

1. White font travels with a dark background. White on a white/transparent surface renders blank — the most common failure. build_html paints its own dark backdrop by default; only go transparent over a host you know is dark. And style both text and foreignObject * with !important, or half the labels stay dark.
2. build_html HTML-escapes the source, so </>/& in node text are safe and round-trip to markmap unchanged — don't pre-escape them yourself.
3. Expansion/search work by editing the Markdown text (regex/filter), not by calling the markmap API. Keep edits scoped to the frontmatter; never stack a second --- block.
4. initialExpandLevel: -1 = expand all.
5. Useful search = match + ancestors + descendants, compared without accents — and ancestry comes from an explicit parent tree, because heading rank and bullet indent share one number line.