Distillary

Agent Retrieval Skill

4 min read

Agent retrieval skill

Any AI agent with web access can query a published Distillary brain. No API keys, no MCP server, no authentication. Just a URL and a skill.

Get the skill

One fetch. Paste it into your agent. Done.

curl https://brain.distillary.xyz/static/skill.txt

Or from GitHub:

curl https://raw.githubusercontent.com/distillary/distillary/main/.claude/skills/distillary-retrieval.md

Every published brain serves its own skill at /static/skill.txt. The agent.json manifest also points to it:

{ "brain": { "skill": "/static/skill.txt" } }

An agent discovering a brain for the first time can fetch agent.json, see the skill URL, fetch the skill, and start querying — fully autonomous, zero human setup.

How it works

  1. Agent fetches {brain_url}/static/agent.json — lightweight manifest with sources, bridges, and navigation instructions
  2. Based on the question type, the agent picks a strategy:
    • Concept lookup → fetch an entity page → backlinks ARE the answer
    • Source summary → thesis is already in the manifest
    • Comparison → fetch the comparison essay or a bridge page
    • Deep evidence → walk the pyramid: root → cluster → structure → atom
    • Exploration → follow backlink chains through the concept graph
  3. Each fetch returns one page with content + links to go deeper
  4. Most questions answered in 2 fetches, under 2000 tokens

Setup by tool

Claude Code

Already built-in. Just say:

Query https://brain.distillary.xyz — what does this brain know about debt?

Codex (OpenAI)

Add to your AGENTS.md or system instructions:

When asked about a Distillary brain, fetch {url}/static/agent.json first.
It tells you what's in the brain and how to navigate it.
Follow links by relevance. Backlinks on entity pages ARE the answers.

Then: codex "Query https://brain.distillary.xyz — what is Barnum's view on luck?"

Gemini CLI

Add to your system prompt or context file:

You can query Distillary brains at any URL.
Start with GET {url}/static/agent.json for the manifest.
Walk pages: entity pages have backlinks = claims about that concept.
2 fetches answers most questions.

Then: gemini "Using https://brain.distillary.xyz — how important is integrity in business?"

OpenCode / Aider / Cline

Add to .opencode/agents.md, .aider.conf.yml, or your system prompt:

# Distillary Brain Retrieval
Fetch {brain_url}/static/agent.json for navigation instructions.
CONCEPT questions → find entity → read backlinks (they ARE the answer).
SOURCE questions → thesis is in manifest → fetch root_note for pyramid.
Every page has: content + wikilinks (deeper) + backlinks (wider).

Cursor / Windsurf / Continue

Add to .cursorrules, .windsurfrules, or config.json system prompt:

When querying a Distillary brain:
1. GET {url}/static/agent.json — returns sources, navigation, bridges
2. For "what is X?" → find entity page → backlinks = claims about X
3. For "summarize" → thesis in manifest → fetch root_note for pyramid
4. For "compare" → fetch bridge concept or /shared/analytics/comparison
5. For "prove it" → walk root → cluster → structure → atom
6. For "explore" → any entity → follow backlinks + wikilinks
7. 2 fetches, <2000 tokens per question

Any agent with HTTP access

The universal skill to paste into any system prompt:

# Distillary Brain Retrieval
 
You can query Distillary brains — published knowledge bases at a URL.
 
## How to use
 
1. Fetch {brain_url}/static/agent.json — returns what's in the brain,
   navigation instructions, and source metadata
2. Pick your strategy based on the question:
   - CONCEPT ("what is X?") → find entity → fetch page → read backlinks
   - SOURCE ("what does this argue?") → thesis in manifest → fetch root_note
   - COMPARISON ("do sources agree?") → fetch bridge concept or /shared/analytics/comparison
   - EVIDENCE ("prove it") → walk pyramid: root → cluster → structure → atom
   - EXPLORATORY ("what's related?") → any entity → follow backlinks + wikilinks
3. Follow 2-3 links for evidence. Most questions: 2 fetches, <2000 tokens.
 
Every page has: content + wikilinks (go deeper) + backlinks (go wider).

That’s it. No SDK. No API key. No MCP server.

Token budget

StrategyFetches~Tokens
Thesis summary1 (agent.json)500
Concept lookup2 (manifest + entity)1500
Cross-source comparison2 (manifest + bridge)2000
Deep evidence chain4 (manifest → root → cluster → atom)2500
Exploration4-6 (follow backlink chains)3000-5000

Graph View

Backlinks