Agentic Hub Docs
Cortex Framework

/distill Workflow

๐Ÿงช /distill Workflow

Standard workflow for extracting structured knowledge from long-form content

๐ŸŽฏ Purpose

Long-form content (ChatGPT journals, Notion pages, exported conversations) can't be directly ingested โ€” they need to be distilled into structured outputs first.

Core idea: Agent reads the full document, extracts categorized outputs, generates a structured change proposal, Adrian reviews and merges. Like a PR for knowledge.

๐Ÿ”„ Flow

Input fileโ†’Agent readsโ†’Extract & classifyโ†’Proposalโ†’Reviewโ†’Merge

๐Ÿ“‹ Step-by-Step Design

Step 1: Input

Agent receives a file path to any long-form content

inputfile path
examplesChatGPT export, Notion page, journal .md

Step 2: Read & Analyze

Agent reads the full document, identifies themes, and scans for extractable items

methodSequential read with view_file, chunked for large files
outputInternal understanding + item candidates

Step 3: Extract & Classify CORE

Categorize extracted items into 5 MECE output types

๐ŸŸฃ PrinciplesNew beliefs, axioms, decision rules โ†’ Principles.md
๐Ÿ”ต FactsNew verified information โ†’ kb/*/facts.md
๐ŸŸข ActionsTodos, tasks to create โ†’ D1 tasks or Smart Todo
๐ŸŸก InsightsPatterns, frameworks, models โ†’ KB articles
๐ŸŸ  ObservationsUnderstanding of Adrian โ†’ adrian.md / agent rules

Step 4: Generate Proposal

Structured change proposal โ€” like a PR for knowledge

per itemtype, affected file, before/after, rationale
conflict checkDoes this contradict existing content?
scope checkIs this local or global? Which nexus node?

Step 5: Adrian Review

Human-in-the-loop governance

acceptApply the change
rejectDon't apply
editModify then apply
deferMark for later review

Step 6: Merge & Mark

Apply approved changes, mark source as processed

mergeWrite approved changes to target files
markTag source file as [distilled]
linkAdd reference back to source in outputs

๐Ÿ“Š Output Type Routing

Output Type

Route To

Review Required?

๐ŸŸฃ Principles

kb/adrian/Principles.mdALWAYS

๐Ÿ”ต Facts

kb/*/facts.md (by nexus node)

USUALLY

๐ŸŸข Actions

D1 Task or Smart Todo

OPTIONAL

๐ŸŸก Insights

KB article or kb/*/patterns.md

USUALLY

๐ŸŸ  Observations

adrian.md or agent rules

ALWAYS

Governance rule: High-layer changes (Principles, Identity, Observations) always require human review. Lower-layer changes (Facts, Actions) can be auto-applied if confidence is high.

๐Ÿงช Prototype Evidence

This workflow was prototyped (manually) in the current session:

Source

Outputs Generated

Time

ChatGPT ้ซ˜ๆ–ฏ็™ฝๅ™ช้Ÿณ่งฃ้‡Š.md
7965 lines

5 new Principles
15 ThirdChoice keywords โ†’ new file
2 Agent Observations
8-layer cognition mapping (insight)
Strategic recommendations (actions)

~2 hours (manual)

Target with automated /distill: 30 minutes per document.

โ“ Open Design Questions

Question

Options

Status

How does agent determine nexus node for routing?

Query nexus API / match by keywords / ask Adrian

OPEN

Where to store distill proposals temporarily?

Task folder / temp file / inline in chat

OPEN

How to handle documents too long for one context window?

Chunked processing / summary-then-detail / multi-pass

OPEN

Should distill create a formal report?

Yes (portal report) / No (just proposal) / Optional

LEANING: optional

Dependency: Nexus needs to exist first?

Yes (Phase 0 first) / No (use workspace as proxy)

DECIDED: Can use workspace as proxy initially

๐Ÿ”— Related

๐Ÿ“– Long-form Flow (B)

The processing pattern that uses /distill

๐Ÿ“ฆ Ingest System

Clara's auto-routing โ€” handles short-form after distill handles long-form

โญ Cortex Docs ยท Back to Cortex Overview

โญ Cortex โ€” Docs

Previous Page

Long-form Flow (B)

Next Page