link-farm

Links I've sent myself — mostly X/Twitter posts and GitHub repos — captured from my phone, enriched automatically, and organized into a concept graph so the collection stays findable as it grows.

890 posts (June 2024 – September 2026), 817 fully enriched, 562 filed into 54 concepts derived from the content rather than a fixed taxonomy.

Start here

I want to…	Read
Browse the collection	ai_links_collection_v3.html — self-contained, no build step, just open it
Understand how it works	CLAUDE.md — the full project context, schema, and curation design decisions
Run it on another machine	SETUP.md — the reconstitution checklist
Know why curation works the way it does	CURATION_DESIGN.md, CURATION_REVIEW.md
See what's planned	ARCHITECTURE_PLAN.md

How it works, briefly

I email links to myself from the iPhone share sheet. A scheduled task runs each weekday morning: it pulls the canonical state from this repo, checks Outlook for new mail, extracts and normalizes the URLs, scrapes each post via Chrome, summarizes and tags it, then runs a pipeline that embeds the content, grows a concept graph from it, assigns each post a single primary home, regenerates the HTML/JSON/Markdown artifacts, and pushes everything back here.

SQLite is the source of truth (db/ai_links.db). The JSON, HTML, and Markdown files are generated — don't hand-edit them; run python3 db/rebuild.py.

Two design choices carry most of the weight, and both are explained at length in CLAUDE.md:

• Keep what I send. Anything pushed through the share sheet was deliberate, so the intake filter drops only true noise (receipts, alerts). Hype-packaged or tangential material gets tagged honestly and kept — the bar is "would I be embarrassed by the tagging", not by the content.
• Automation attaches and labels; it never discards. Edges carry a role. evidence/origin are load-bearing and define what a concept means; weak is a recorded association that deliberately doesn't vote. Only a human dismisses anything.

Layout

posts_final_v3.json            # generated dataset
ai_links_collection_v3.html    # generated viewer (open this)
ai_links_collection_v3.md      # generated markdown companion
db/
  ai_links.db                  # SQLite — source of truth
  pipeline.py                  # the post-enrichment pipeline
  enrich.py                    # canonical write path (never bypass it)
  concepts.py  embeddings.py   # the concept graph + semantic layer
  repo_sync.py                 # PUSH_SPEC — what gets mirrored here
  stats.py                     # regenerates the stats block in CLAUDE.md
  test_*.py                    # roles / lock / recover
scheduled/                     # snapshots of the live scheduled tasks
skills/                        # skills that travel with the repo

Quick checks

pip install --break-system-packages -r requirements.txt
python3 db/migrate_runner.py      # schema up to date
python3 -m db.stats               # corpus + concept graph summary
python3 -m unittest discover -s db -p 'test_*.py'

db/test_roles.py is the one that matters most — it guards the four places where the canonical/weak edge distinction is enforced, each of which fails silently if it regresses.