# Research Command

> 1 Prompt → Interconnected Knowledge: By giving Claude the domain topic and specific formatting rules, you can generate a fully connected web of knowledge with real backlinks from one prompt.

To use: /research about post training quantization in gemma model for local inference

*Prompt · May 31, 2026*

## description: Deep research a topic, write a sourced note in the Obsidian brain vault, and fix backlinks

Research the topic provided and create a well-sourced note in the user's Obsidian vault at ⁠ Custom Path ⁠.

Topic: $ARGUMENTS

## Phase 1: Research

Do thorough research using multiple real sources. Use web search, documentation tools (context7), and any other grounded tools available. Aim for 3-6 distinct, high-quality sources.

For each source, record: •⁠ ⁠The URL •⁠ ⁠The key claims/facts you extracted from it

Do NOT use training data as a source. Every fact in the final note must trace back to a URL you actually visited and read during this session.

## Phase 2: Write the Note

Write a markdown file in ⁠ Custom Path ⁠. Use this frontmatter format:

## ⁠ yaml

title: "Topic Title" sources:

- url: "https://actual-url-you-read.com" title: "Page title or description"
- url: "https://another-source.com" title: "Page title or description" created: YYYY-MM-DD description: "One-line description" tags:
- "research"

---

⁠

Writing rules: •⁠ ⁠Write in the user's style: concise, informal, focused on the "why" not just the "what". Look at existing notes in ⁠ Custom Path ⁠ for tone reference. •⁠ ⁠Use ⁠ \[\[double brackets\]\] ⁠ for any concept that has or should have its own note in the vault. •⁠ ⁠Inline source references where claims are made, like: "Flash Attention reduces memory from O(N^2) to O(N) by tiling ([source](url))". Don't just dump sources in frontmatter -- connect them to specific claims. •⁠ ⁠Prefer bullet points and short paragraphs over walls of text. •⁠ ⁠Focus on building understanding, not being comprehensive. What are the key ideas someone needs to grasp?

## Phase 3: Backlink Scan

After writing the new note:

1.⁠ ⁠*Forward links*: You already added ⁠ \[\[backlinks\]\] ⁠ in the note during writing. Verify each linked concept actually exists as a file in the vault. If it doesn't exist, that's fine -- Obsidian handles dangling links. Just make sure the names match existing files where they exist (check with glob).

2.⁠ ⁠*Backward links*: Scan all existing ⁠ .md ⁠ files in ⁠ Custom Path ⁠ (not in subdirectories like Clippings/). For each file, check if the new topic is mentioned or closely related. If a natural backlink opportunity exists:

- Read the existing note
- Add a ⁠ \[\[New Topic\]\] ⁠ link where it contextually fits (not forced)
- Only edit if the link genuinely adds value -- don't shoehorn links into every file

## Phase 4: Report

Show the user: •⁠ ⁠The note that was created (file path) •⁠ ⁠Sources used (title + URL for each) •⁠ ⁠Any backward links added to existing notes •⁠ ⁠Any ⁠ \[\[concepts\]\] ⁠ referenced that don't have notes yet (potential future research topics)

## Rules

•⁠ ⁠NEVER fabricate a URL. Every source must be a real page you fetched and read. •⁠ ⁠NEVER write claims you can't trace to a specific source from this session. •⁠ ⁠If you can't find good sources on a topic, say so -- don't fill the gap with training data. •⁠ ⁠Ask the user before editing more than 3 existing files for backward links. •⁠ ⁠File names should be Title Case with spaces (matching the existing vault convention, e.g., ⁠ Flash Attention.md ⁠)

**Tags:** wiki, obsidian, research

[Open on the site →](https://www.agenticamit.com/prompt/research-command)