LLM Content Block

English · 中文

Paste AI conversations into Obsidian as a scrollable, format-safe card.

Why

We're entering a phase where more of what we want to keep is conversations with AI — a Q&A, a brainstorm, a debugging session. These conversations are themselves worth saving, the way we used to save articles, web pages, book notes.

But pasting them straight into Obsidian breaks immediately:

It wrecks your formatting. An AI reply is itself markdown — headings, lists, code blocks, math. Once pasted, Obsidian treats it as your note's own structure: a ## heading inside the reply becomes a level-2 heading in your note, its lists merge into your lists, its ``` fences can even cut off your other blocks. Paste a few conversations and the whole note's hierarchy is scrambled — you can no longer tell what's your note and what's the AI's answer.
It doesn't fit. A good AI reply is often hundreds of words; it fills the screen and stretches the note endlessly.
No structure. A conversation is back-and-forth turns; flattened into one blob, the dialogue is lost.

LLM Content Block gives AI conversations their own container, fixing all three:

The conversation lives inside a card; its markdown renders only inside the card and never leaks into the note body — the AI's headings don't become your headings, its code blocks don't slice your note. Your note stays clean.
The card has a max height and scrolls internally; long AI answers even collapse by default — click Expand for the full text.
Turns are recognized as user / AI bubbles, with a source tag (GPT / Claude / …).

In one line: make AI-generated content a self-contained, collectible, format-safe block in Obsidian — like a callout is a container for emphasis, this is a container for AI conversations.

Usage

Write a fenced code block with language aicard and paste the conversation:

```aicard
source: Claude
You: What's the difference between RSC and SSR?
Claude: The essence is **where it renders and when it hydrates**… (markdown, code, LaTeX all fine)
```

source: (optional): the model — click the pill at the card's top-right to switch or type any name (presets: GPT / Gemini / Claude / DeepSeek / Qwen / ERNIE / Kimi / Doubao / Gemma / GLM / Grok / Llama, plus custom).
Lines starting with Role: split into left/right bubbles: me/you/user → right, Claude/GPT/AI/assistant → left; other roles get a neutral bubble.
Plain text with no role prefix becomes one clean content card.
Full markdown support: code blocks, tables, LaTeX ( $...$ / $$...$$), callouts, wiki links.
Edit button (top-right) edits the conversation and source in place — the edit box is plain text, so pasting won't scramble formatting.

If the conversation itself contains ``` code blocks, use 4 backticks ````aicard for the outer fence to avoid truncation. There's also a command Insert AI chat block: with text selected → wrap it into a card; no selection → insert a skeleton. Bind a hotkey in Settings → Hotkeys.

Features

Format isolation: the AI content's markdown is contained in the card and never pollutes your note.
Scroll + collapse: the card is height-limited and scrolls; long AI answers collapse to a preview.
Rich text: code highlighting, LaTeX, tables, callouts, wiki links.
Dialogue structure: auto user/AI bubbles; recognizes Chinese & English role names.
Source management: 12 preset models + custom; one-click switch via the pill, written back to the note.
Theme-aware: built entirely on Obsidian CSS variables; follows dark/light automatically.
i18n: UI follows Obsidian's current language (中文 / English).

Settings

Card max height, body font size, body font family (empty = follow theme).
Custom sources (name + optional brand color).

View

Reading view renders perfectly. In Live Preview, very long blocks may need scrolling near the block's end to trigger rendering — for archived content, read in Reading view for the best experience.

Development

npm install
npm run build          # type-check + bundle
npm run build:deploy   # bundle + deploy to vault
npm run dev            # watch

The deploy target is set via env var or a local file (the repo itself contains no personal path):

export OBSIDIAN_VAULT=/path/to/your/vault   # option 1
echo /path/to/your/vault > .deploy-vault     # option 2 (.deploy-vault is gitignored)

License

MIT License — see LICENSE.