Content Pipeline

1. What? — Definition and context

An idea hits you in the shower — you record a voice note on Telegram. An interesting article shows up in your feed — you forward the URL to the bot. You snap a photo of a whiteboard after a meeting. Each time, the Content Pipeline structures the content with Claude and stores it as an Obsidian note inside a GitHub vault — ready to be retrieved, expanded, or published.

Obsidian

Obsidian is a Markdown note editor that keeps everything as local files. By using a GitHub repository as the vault, notes are versioned, accessible everywhere, and synchronised across devices.

Six content types supported

Command	Type	Input	Output
/note	Note	Free text	Structured note with wikilinks
/blog	Blog	Free text	Article with intro, body, conclusion
/research	Research	Topic	In-depth analysis with sources
(URL)	URL	Web link	Summary + key points + citations
(voice)	Voice	Audio	Cleaned, structured transcript
(photo)	Photo	Image	Corrected, structured OCR text

Architecture

---

2. Why? — Stakes and motivations

Problems solved

Problem	Without pipeline	With pipeline
Lost ideas	Note in some corner, never found	Indexed and versioned Obsidian vault
Formatting	Raw text without structure	Claude structures with frontmatter
Scattered content	Notes here, bookmarks there	One vault, one workflow
Painful publishing	Copy-paste between apps	/blog → ready-to-publish article

Why a review workflow?

Claude-generated content is not stored straight away. The user first gets a Telegram preview with the title, tags, target path and the first 500 characters. Two buttons let you confirm or cancel. It is a safety net against bad classification or malformed notes.

Automatic expiration

Pending content expires after 24 hours. A cleanup workflow removes unconfirmed entries to avoid build-up.

---

3. How? — Technical implementation

The journey of a piece of content

1. Reception — The Telegram Orchestrator detects a /note, /blog, /research command, a URL, a voice message or a photo and routes it to the Content Pipeline.

2. Consolidation — A Code node normalises the input: text is extracted, URLs are fetched (truncated to 8000 characters), voice messages arrive already transcribed, photos already OCR’d.

3. Claude generation — The LLM (Claude Sonnet via CLI Ollama, 120 s timeout) receives a content-type-specific prompt. Each prompt defines the expected structure, the YAML frontmatter, and the formatting conventions (wikilinks for notes, sections for blogs, sources for research).

4. Temporary storage — The generated Markdown lands in the n8n_pending_actions Data Table with a short identifier (8 hex chars) and a 24 h expiration.

5. Preview — The user gets a Telegram message with the content preview and two buttons:

📝 New note

Title: Docker Networking Guide
Tags: docker, networking, guide
Path: notes/tech/2026-02-19-docker-networking-guide.md

Docker networking in practice involves understanding
bridge networks, overlay networks, and host mode...
[... truncated]

[Save] [Cancel]

6. Save — Clicking [Save] tells the Orchestrator to create the file in the guigpap/obsidian-vault GitHub repo via the API and send a confirmation.

Automatic classification

The storage path is decided by type and tags:

Condition	Path
Type = research	knowledge/
Type = blog	notes/projects/
Tag contains “project”	notes/projects/
Tag contains “reference”	notes/reference/
Default	notes/tech/

The file is named {date}-{slug}.md (e.g. 2026-02-19-docker-networking-guide.md).

Output format

Every note is a Markdown file with YAML frontmatter:

---
title: "Docker Networking Guide"
tags: [docker, networking, guide]
date: 2026-02-19
source: "telegram"
status: "draft"
type: "note"
slug: "docker-networking-guide"
---

For blog articles, two extra fields are produced: description (meta description) and a structure with Introduction, Body, Conclusion.

For research items, the body includes: Overview, Key points, Technical details, Sources, Related topics.

---

4. What if? — Outlook and limits

Current limits

Limit	Impact	Mitigation
No deduplication	Sending the same URL twice = two notes	The user can cancel
Imperfect OCR	Low-quality photos → messy text	Claude fixes obvious mistakes
No vault search	Finding a note = browsing files	Tags + Obsidian search + Qdrant (planned)

Evolution scenarios

If automatic publishing is needed:

• The Obsidian Blog Publisher workflow can take over
• Blog notes flagged status: ready are publishing candidates

If semantic search is needed:

• Generate embeddings for every note
• Store them in Qdrant for vector search
• /search Telegram command to query the vault

If note volume grows:

• Add sub-folders by month or project
• Keep an automatic Obsidian index
• Weekly digest of recent notes

---

Related pages

Workflows

• Telegram Orchestrator — Routing of content commands
• Voice Transcription — Transcription of voice messages
• YouTube Digest — Similar pipeline for videos

Infrastructure

• AI Stack — CLI Ollama for Claude