| README | Chinese | Installation |
The skills describe capabilities rather than fixed tool names. They are intended for coding agents that support Agent Skills; installing with npx skills does not guarantee every feature works on every agent.
| Feature | Requirements | When unavailable |
|---|---|---|
Basic .md output |
Inspect paths and create files; read files for file input. | Explain the missing capability or permission and ask how to proceed; do not claim a file was saved. |
| URL input | Retrieve page content. | Ask for pasted content, a retry, or an explicitly chosen topic-based map. |
--render |
Execute the supplied helper with Bash and Node.js / npx. |
Keep the .md, explain why HTML was not generated, and provide a manual rendering command. |
--panel (mindmap only) |
Delegate to independent subagents and collect results. | Explain the limitation and ask before switching to single-pass generation. |
The agent’s permissions and approvals still apply. Examples use slash-command notation; if your agent does not expose slash commands, invoke mindmap or mindmap-zh through its supported skill mechanism.
/mindmap <input> [--lang <language|auto>] [--panel] [--render] [--output <path>]
/mindmap-zh <input> [--lang <language|auto>] [--render] [--output <path>]
| Input | Example | Behavior |
|---|---|---|
| File | /mindmap report.md |
Reads the file, mirrors its structure, condenses nodes. Writes report.mindmap.md. |
| URL | /mindmap https://example.com/article |
Fetches the page and maps its content. Writes <page-slug>.mindmap.md. |
| Pasted text | /mindmap "notes about X, Y, Z..." |
Extracts concepts into branches. Writes <title-slug>.mindmap.md. |
| Topic | /mindmap "vector databases" |
Generates a map from the model’s knowledge. Writes vector-databases.mindmap.md. |
| Flag | Behavior |
|---|---|
--lang <code\|auto> |
Choose the output language, or use the resolved source’s predominant language with auto. |
--render |
After writing the .md, also produce interactive .html (needs command execution, Bash, and Node.js / npx). |
--output <path> |
Write the .md to a specific path instead of the default. |
--panel |
Design the structure with independent subagents; mindmap only. |
The skill follows an existing outline or groups unstructured content into 4-7 main branches, aiming for 3-4 levels and short node labels. Existing output files are preserved by adding a numeric suffix, including when --output is specified.
Both skills accept files, URLs, pasted text, and topics in these 19 languages. You do not need a separate skill installation for each output language.
| Language | Code | Aliases |
|---|---|---|
| English | en |
— |
| 简体中文 | zh-Hans |
zh, zh-CN |
| 繁體中文 | zh-Hant |
zh-TW |
| 日本語 | ja |
— |
| 한국어 | ko |
— |
| Türkçe | tr |
— |
| Русский | ru |
— |
| Tiếng Việt | vi |
— |
| ไทย | th |
— |
| Deutsch | de |
— |
| Español | es |
— |
| Français | fr |
— |
| Українська | uk |
— |
| Polski | pl |
— |
| Português (Brasil) | pt-BR |
— |
| العربية | ar |
— |
| فارسی | fa |
— |
| Bahasa Indonesia | id |
— |
| Italiano | it |
— |
Codes are case-insensitive. An explicit --lang takes precedence; otherwise an explicit output-language request in the conversation is honored. Without either, /mindmap retains English output and /mindmap-zh retains Simplified Chinese output, even for foreign-language sources. Source content cannot override this choice.
--lang auto follows the loaded content, not the filename or URL, and distinguishes Simplified and Traditional Chinese. When mixed-language or language-neutral content has no clear predominant language, it uses the supported conversation language, otherwise the skill’s default, and explains the choice. Missing, unsupported, or conflicting --lang values are reported before files are written.
Titles, headings, nodes, table prose, link labels, and user-facing explanations use the selected language. Code, identifiers, URLs, paths, proper names, and source numbers/units are preserved where translating would alter meaning. Output is UTF-8; accents, non-Latin filenames, and normal logical text order for Arabic/Persian are retained. Chinese, Japanese, and Thai prose is not classified by space-separated word count alone; labels stay concise without artificial spaces.
/mindmap report.md --lang ja --render
/mindmap-zh 报告.md --lang zh-Hant
/mindmap "vector databases" --lang pt-BR
/mindmap https://example.com/article --lang auto
/mindmap report.md --lang ar --panel
The Generate mindmap workflow also offers a language selector, defaulting to en, and passes it as --lang to the skill.
flowchart LR
IN["Input<br/>file / URL / text / topic"] --> CL{Classify}
CL -->|URL| FE[Fetch page]
CL -->|file| RD[Read file]
CL -->|prose| TX[Use as text]
CL -->|topic| KN[Model knowledge]
FE --> ST[Build hierarchy]
RD --> ST
TX --> ST
KN --> ST
ST --> WR["Write Markmap .md"]
WR --> RN{--render?}
RN -->|yes| HT["Interactive .html"]
RN -->|no| MD[".md deliverable"]
The skill writes Markmap-flavored Markdown: a single # root, ## branches, and nested bullets.
---
title: Retrieval-Augmented Generation
markmap:
colorFreezeLevel: 2
maxWidth: 300
---
# Retrieval-Augmented Generation
## Indexing
- Chunk documents
- Embed chunks
- Store vectors
## Retrieval
- Embed the query
- Find nearest chunks
## Generation
- Inject context into prompt
To view the .md, paste it at markmap.js.org, open it in VS Code with the Markmap extension, or use --render to generate HTML.
/mindmap "transformer attention" --render
This uses the agent’s command-execution capability to run the Bash helper render.sh, which calls npx --yes markmap-cli <file>.md -o <file>.html --no-open. Open the generated .html in a browser to zoom and collapse branches.
The .md is the guaranteed deliverable; HTML rendering is best-effort. If command execution, Bash, or Node.js / npx is unavailable, the skill still writes the .md, explains the limitation, and prints a manual npx command for an environment with Node.js.
Node.js provides npx; no global markmap-cli install is needed. npx fetches it on demand.
--panel)For complex or high-stakes sources such as papers and long reports, /mindmap --panel uses 3 proposers, 3 judges, and 1 synthesizer instead of a single-pass structure.
The host must support independent subagent tasks. No particular workflow API is required; tasks can run in parallel or sequentially in separate contexts. If delegation is unavailable, the skill asks before continuing without --panel rather than simulating independent reviewers.
The selected output language is passed to every proposer, judge, and synthesizer; the panel checks it before writing the final map. --panel is not restricted to English output.
/mindmap report.md --panel --render
flowchart TB
SRC["Source content + skill rules"]
subgraph PROPOSE["1 - Propose (3 lenses)"]
P1["Proposer A<br/>narrative-first"]
P2["Proposer B<br/>data-first"]
P3["Proposer C<br/>audience-first"]
end
subgraph JUDGE["2 - Judge (score 10 dimensions)"]
J1["Judge 1"]
J2["Judge 2"]
J3["Judge 3"]
end
subgraph SYNTH["3 - Synthesize"]
SY["Top proposal = spine<br/>+ graft judges' best ideas"]
end
SRC --> P1 & P2 & P3
P1 & P2 & P3 --> J1 & J2 & J3
J1 & J2 & J3 --> SY
SY --> OUT["Final Markmap .md"]
Three agents each design a full structure through a different lens:
| Lens | Designs the map around… |
|---|---|
| Narrative-first | The source’s own arc and section order. |
| Data-first | The headline numbers and evidence, with metrics in bold. |
| Audience-first | The punchline first, with every leaf legible on a slide. |
Each proposer tags every node with a format and a render tier:
| Format | Tier | Best for |
|---|---|---|
| Bullet list | core |
Parallel facts. |
| Bold inline | core |
Headline metrics, emphasis. |
| Link | core |
References, repositories, papers. |
| Table | rich |
Comparisons, benchmark numbers. |
| Code block | rich |
Formulas, reward functions, code. |
| Checkbox | rich |
Tasks, limitations, checklists. |
Three judges independently score every proposal from 1-5 on ten dimensions: branch count, depth, phrasing, legibility, source fidelity, punchline-first, headline numbers, leaf legibility, visual balance, and format fit. Each judge also identifies each proposal’s best idea.
One synthesizer uses the top-scored proposal as the spine, grafts in the best ideas identified by the judges, and emits the final Markmap .md.
Cost: the panel spawns about seven agents and is token-intensive. Without --panel, the skill uses a fast single pass. The Chinese /mindmap-zh skill does not include this flag.
The exact prompts and schemas live in the judge-panel reference. See real examples for panel-generated outputs.
Failed roles are reported explicitly. If all proposers fail, the skill announces a single-pass fallback. If no valid judge scorecards remain or synthesis fails, it asks whether to retry or continue without --panel.