Imported from Talegorithm/Glynk (SKILL.md). Install upstream with npx skills add Talegorithm/Glynk. Copyright stays with the author.

Glynk Content Platform

Glynk ingests long-form content (books, papers, podcasts, articles), performs structured processing, and exposes it for annotation and semantic search by humans and agents.

Prerequisites

First check if environment variables are already set:

echo "GLYNK_TOKEN=${GLYNK_TOKEN:-(not set)}" && echo "GLYNK_API_URL=${GLYNK_API_URL:-(not set)}"

If set, skip to the next section. Otherwise, register:

# 1. Try to find user's email (git config or environment)
git config user.email  # or echo $EMAIL

# 2. Ask user for uid and email (leave email null if no reliable source found)
# 3. Register
curl -X POST $GLYNK_API_URL/api/users \
  -H "Content-Type: application/json" \
  -d '{"uid":"user-specified-uid","email":"user-email-or-null"}'
# -> {"uid":"...","token":"glk_..."}

# 4. Set environment variables
export GLYNK_TOKEN="glk_..."
export GLYNK_API_URL="https://brainow.link"  # public frontend origin; /api is reverse-proxied to backend

Note: both uid and email are optional. uid is auto-generated if omitted. Never fabricate an email -- leave it null if not found.

All endpoints below require Authorization: Bearer $GLYNK_TOKEN unless noted.

Core Workflow

Every piece of content on Glynk is a Unit, but there are two kinds of publishing:

publication (shape=structured): readable, span-annotatable content (books, articles, transcripts, md uploads). Created via POST /api/publications or POST /api/publications/upload.
thought (shape=flat): a standalone authored utterance (comment, note, reaction). Created via POST /api/thoughts.

Read / list / search / annotate APIs all operate on Unit regardless of shape.

List: GET /api/units?origin=ingested&limit=50 -- browse publications
Read: GET /api/units/{id}/read?size=20000 -- read page by page (publications paginate by spans; thoughts return single-shot)
Annotate: POST /api/anchors or /api/anchors/batch -- create highlights/hooks/notes/etc.
Search: POST /api/units/search -- semantic search across all Units
Outline: PUT /api/units/{id}/outline -- submit structured outline

Key Concepts

Span IDs

Every sentence in content has a unique span ID: {content_id}-{file_idx}-p{paragraph}-s{sentence}

Example: a1b2c3d4-0-p15-s3 = content a1b2c3d4, file 0, paragraph 15, sentence 3.

Span IDs are embedded in HTML as <span id="..."> and used as anchors for all annotations.

Anchor Roles

See glynk/models.py ROLE_SCHEMAS for the authoritative source (enforced at API).

role	source	target	body	Purpose
`highlight`	unit	span	auto (= span text)	Text selection with color
`hook`	unit	span	required	Agent-raised question on a span
`note`	unit	span \| unit	required	Free-form note
`summary`	unit	unit	required	TL;DR of a whole Unit
`reply`	unit	span \| unit	optional	Threaded discussion (emoji / image / text)
`like`	entity	span \| unit	none	Lightweight endorsement
`bookmark`	entity	span \| unit	none	Personal save
`follow`	entity	entity	none	Subscribe to an entity

Anchor Format

{
  "type": "text",
  "spans": ["a1b2-0-p1-s1", "a1b2-0-p1-s2"],
  "color": "ghost"
}

Colors: yellow, green, blue, pink, ghost (invisible highlight, used by agents).

Hooks (Agent Annotations)

Hooks are reverse-engineered questions: assume the content is the "answer" -- what question would lead someone here?

Decontextualized: understandable without reading the source
Precise: point to 1-N consecutive spans that directly answer the question
Thought-provoking: dialectical or insightful, not information restatement
Tagged: 2-5 abstract keywords for discoverability

Good: "How do you make decisions with incomplete information?" Bad: "What does Peter Thiel think about secrets?"

API Reference

Content

# Get content metadata + TOC + outline
curl "$GLYNK_API_URL/api/content/{content_id}"

# Ingest from URL
curl -X POST "$GLYNK_API_URL/api/publications" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"source":"https://example.com/article"}'
# -> {"content_id":"a1b2c3d4","title":"...","source_type":"article","file_count":1,"total_chars":5000}

# Upload file directly
curl -X POST "$GLYNK_API_URL/api/publications/upload" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -F "file=@book.epub"

Reading Content

# Read a chunk (AI view -- simplified HTML, token-efficient)
curl "$GLYNK_API_URL/api/content/{content_id}/chunk?size=20000"
# -> {"content":"<html>...<span id='a1b2-0-p1-s1'>text</span>...</html>",
#    "from":"a1b2-0-p1-s1","to":"a1b2-0-p3-s5","char_count":17945,
#    "has_more":true,"next_from":"a1b2-0-p4-s1"}

# Continue reading from where you left off
curl "$GLYNK_API_URL/api/content/{content_id}/chunk?size=8000&from=a1b2-0-p4-s1"

# Read full file (human view -- complete HTML with styling)
curl "$GLYNK_API_URL/api/content/{content_id}/file"

Pagination: use next_from from each response as the from parameter for the next request. Continue until has_more is false.

Annotations

# Create single annotation
curl -X POST "$GLYNK_API_URL/api/annotate" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "content_id": "a1b2c3d4",
    "anchor": {"type":"text","spans":["a1b2-0-p5-s1","a1b2-0-p5-s2"],"color":"ghost"},
    "type": "hook",
    "text": "How do you make decisions with incomplete information?",
    "tags": ["decision-making", "uncertainty"],
    "contextuality": "standalone"
  }'

# Batch create (preferred for agents)
curl -X POST "$GLYNK_API_URL/api/annotate/batch" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"annotations":[...]}'
# -> {"created":27,"ids":["ann-1","ann-2",...]}

# List my annotations
curl "$GLYNK_API_URL/api/annotations?content_id=a1b2c3d4&type=hook&limit=50" \
  -H "Authorization: Bearer $GLYNK_TOKEN"
# -> {"annotations":[...],"total":150}

# Update annotation
curl -X PATCH "$GLYNK_API_URL/api/annotations/{annotation_id}" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"text":"Updated question"}'

# Delete annotation
curl -X DELETE "$GLYNK_API_URL/api/annotations/{annotation_id}" \
  -H "Authorization: Bearer $GLYNK_TOKEN"

Semantic Search

# Search across all annotations
curl -X POST "$GLYNK_API_URL/api/query" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "How to make decisions with incomplete information",
    "types": ["hook", "highlight"],
    "top_k": 10
  }'
# -> {"query_id":"qry-...","results":[
#     {"id":"ann-123","type":"hook","text":"...","spans":[...],"score":0.92,"content_id":"a1b2c3d4"}
#   ]}

# Search only my annotations
curl -X POST "$GLYNK_API_URL/api/annotations/search" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"query":"decision making under uncertainty"}'

# Submit feedback on search results (improves ranking)
curl -X POST "$GLYNK_API_URL/api/feedback" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"query_id":"qry-...","results":[{"result_id":"ann-123","clicked_through":true}]}'

Outlines

Outlines should be submitted after reading the entire content, not incrementally (unlike hooks). Maintain a local temporary JSON file while reading, updating the outline structure as you go, then submit the complete version once finished.

Outline Structure Requirements

Each entry contains:

title: Short, powerful title (5-15 characters)
description: 1-2 sentence description -- reading top-to-bottom should feel coherent, like a simplified version of the original
span_id: Starting position in content (extracted from <span id="..."> in HTML)
children: Sub-entries (recursive)

Requirements:

Natural hierarchy, no forced depth, 2-5 items per level
Faithful to original content, no invented information
Full-length book: 10-30 top-level items, depending on content length

# Submit AI outline (overwrites existing, submit after reading all content)
curl -X PUT "$GLYNK_API_URL/api/content/{content_id}/outline" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"outline":[
    {"title":"Zero to One Thinking",
     "description":"Innovation is not copying from 1 to N, but creating from 0 to 1. Every important moment is unique.",
     "span_id":"a1b2-0-p1-s1",
     "children":[
       {"title":"Contrarian Questions","description":"Great startups begin with a truth most people disagree with.","span_id":"a1b2-0-p3-s1","children":[]},
       {"title":"Monopoly vs Competition","description":"Competition is for losers; real profits come from monopoly.","span_id":"a1b2-0-p8-s1","children":[]}
     ]}
  ]}'

# Get outline
curl "$GLYNK_API_URL/api/content/{content_id}/outline"

RSS Sources

# Add RSS feed for auto-ingestion
curl -X POST "$GLYNK_API_URL/api/sources" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"url":"https://example.com/feed.xml","schedule":"daily","max_items":5}'

# List / update / delete sources
curl "$GLYNK_API_URL/api/sources" -H "Authorization: Bearer $GLYNK_TOKEN"
curl -X PUT "$GLYNK_API_URL/api/sources/{id}" ...
curl -X DELETE "$GLYNK_API_URL/api/sources/{id}" ...

Common Agent Workflows

Read and Annotate Content

User specifies a content (or provides a file/URL to ingest first), then read page by page, generating hooks and outline.

# 1. Ingest content (if user provided a file/URL)
curl -X POST "$GLYNK_API_URL/api/publications" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"source":"https://example.com/article"}'

# 2. Read page by page, collecting span IDs
curl "$GLYNK_API_URL/api/content/a1b2c3d4/chunk?size=20000"
# ... use next_from to paginate ...

# 3. Submit hooks as you read (batch every ~20 hooks)
curl -X POST "$GLYNK_API_URL/api/annotate/batch" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"annotations":[
    {"content_id":"a1b2c3d4","type":"hook",
     "anchor":{"type":"text","spans":["a1b2-0-p5-s1"],"color":"ghost"},
     "text":"How do you identify opportunities others miss?",
     "tags":["opportunity","contrarian-thinking"],"contextuality":"standalone"}
  ]}'

# 4. Submit outline after reading all content
curl -X PUT "$GLYNK_API_URL/api/content/a1b2c3d4/outline" ...

Search for Relevant Knowledge

# Semantic search -- finds hooks/highlights across all content
curl -X POST "$GLYNK_API_URL/api/query" \
  -H "Content-Type: application/json" \
  -d '{"text":"strategies for dealing with uncertainty","top_k":5}'

# Read the original context around a result
curl "$GLYNK_API_URL/api/content/{content_id}/chunk?size=3000&from={span_id}"

Tips

Use chunk endpoint for reading -- returns simplified HTML optimized for LLM token efficiency
Batch annotations -- use /annotate/batch instead of individual calls
Track reading with next_from -- each chunk response tells you where to continue
Agents use ghost color -- agent annotations don't show visible highlights to readers
Hooks enable discovery -- write hooks as questions someone might search for, not as summaries
No special privileges -- agents use the same public API as any other client

When the User Gives You a URL or File Path

Step 1: Ingest Content

# URL
curl -X POST "$GLYNK_API_URL/api/publications" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"source":"the-url-from-user"}'

# Local file
curl -X POST "$GLYNK_API_URL/api/publications/upload" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -F "file=@path-to-file"

After successful ingestion you get a content_id. All subsequent operations use this ID.

Step 2: Read the Entire Content

Use the chunk endpoint to read page by page (AI view):

curl "$GLYNK_API_URL/api/content/{content_id}/chunk?size=20000"
# Use next_from to paginate until has_more is false

While reading, do two things:

2.1 Build an Outline

Maintain a local outline structure. Update it as you read, submit once after finishing the entire content:

curl -X PUT "$GLYNK_API_URL/api/content/{content_id}/outline" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"outline": [...]}'

See the "Outlines" section above for structure requirements.

2.2 Annotate: Highlights / Hooks

As you read, select the most valuable passages for the user and create two types of annotations:

highlight: Passages worth remembering (brilliant arguments, key evidence, counterintuitive insights)
hook: Reverse-engineered questions (see "Hooks" section above)

Principles:

Position: Choose passages the user would want to highlight themselves, but write annotation text that is valuable to any reader -- don't write it as a personal conversation or include too much user-specific content (save that for the report in Step 3)
Length: Each annotated passage should be a self-contained unit of meaning -- a full argument paragraph or a standalone one-liner, but always independently readable
Quantity: Quality over quantity. Pick the most valuable, most distinctive passages. Unless the user asks otherwise, roughly 2-5 annotations per 10,000 characters is appropriate.

curl -X POST "$GLYNK_API_URL/api/annotate/batch" \
  -H "Authorization: Bearer $GLYNK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"annotations":[
    {"content_id":"...","type":"highlight",
     "anchor":{"type":"text","spans":["xxx-0-p12-s1","xxx-0-p12-s2"],"color":"ghost"},
     "text":"Original text (used for embedding search)",
     "tags":["keyword"],"contextuality":"standalone"},
    {"content_id":"...","type":"hook",
     "anchor":{"type":"text","spans":["xxx-0-p12-s1","xxx-0-p12-s2"],"color":"ghost"},
     "text":"How do you make decisions with incomplete information?",
     "tags":["decision-making","uncertainty"],"contextuality":"standalone"}
  ]}'

Step 3: Report to the User (Core Deliverable)

The report is your most important output. The user wants to genuinely understand the content through your report, not just see a list of fragmented annotations.

Report Structure

Summary: One paragraph capturing the core ideas and the author's stance
Key Takeaways (main body): Select the 3-5 most important points that are most relevant to the user. The selection criterion is: if the user read this themselves, which passages would they be moved by? What would they highlight? Each takeaway includes:
- Your commentary: Why this point matters to the user. Connect it to their situation, interests, or current work. This part can and should be personalized.
- Original text: Quote the original text generously (not just one or two sentences, but a full argument paragraph), so the user can read the author's own words. Use blockquote (>) format.
- Link: [Read in context ->](https://brainow.link/read/{content_id}?loc={span_id}) Takeaways should flow logically from one to the next, reading like a coherent guided tour rather than isolated excerpts.
Overall reflection (optional): If you have a big-picture observation about the content (e.g., how it connects to the user's current focus), briefly share it at the end.

Link Format

https://brainow.link/read/{content_id}?loc={span_id}

Example: https://brainow.link/read/a1b2c3d4?loc=a1b2c3d4-5-p12-s1

Clicking the link opens the reader and scrolls to the corresponding paragraph. (User needs to log in with their token in the browser first.)

glynk

About