AI is a powerful tool, and I wanted to learn to use it beyond my everyday design work. So I tried building software (a small enter-the-Matrix moment). I spend more time on Instagram than I'd like to admit, so I decided to turn some of it into something I value. Whenever an Instagram video inspires me, a custom iOS shortcut sends it straight into my own visual library. How it all works is still a bit of a mystery to me, but it works.
The Instagram URL is captured and added to a queue. From there, a small app I built downloads the video, pulls the creator information from the post, uploads the video to Cloudflare, and keeps track of the process through Supabase. The original creator is always credited!
The result is my own visual inspiration library, shaped by what I find interesting: nature, spaces, objects, transportation, art and the incredible things people make. Being online is part of being a creative, and so are algorithms. But I don't want an algorithm shaping my taste. I'd rather curate it myself. At the other end of every algorithm are amazing people, ideas and creations. The algorithm should help me find them, not decide what I care about.
I'm not anti-technology, anti-AI or anti-social media. These tools are here, and they can be really useful. What matters is using them on purpose, so we control the tools instead of the other way around. In a small way, this project does exactly that: it turns scrolling, which can easily become mindless, into a source of intentional inspiration.
01 · Capture — iOS Shortcut / browser
Shortcut “Humane” (share sheet)
Accepts URLs + Text from any app. Normalizes the link — regex replace \?.* → “” strips ?igsh=… tracking so dedup keys on the canonical URL.
Then a raw REST call, no SDK: POST /rest/v1/inbox_links?on_conflict=url — headers apikey: <anon> · Authorization: Bearer <anon> · Content-Type: application/json · Prefer: resolution=ignore-duplicates — body {“url”: <cleaned link>}. Re-sharing the same reel is a silent no-op.
/add page (fallback + monitor)
Static HTML in a Squarespace code block, anon key baked in. Same insert via fetch(), plus GET …?status=eq.pending&order=created_at.desc to render the live queue — doubles as a health check without touching the PC.
The anon key is public by design — RLS below is what makes that safe
02 · Queue — Supabase (Postgres + PostgREST)
The queue is the only state in the system
create table inbox_links (
id bigint generated always as identity primary key,
url text unique not null, -- dedup constraint the upsert targets
status text default ‘pending’, -- pending → done; failures stay pending
created_at timestamptz default now() );
Row-level security on. Two policies for anon: INSERT and SELECT only — the public key physically cannot update or delete. The service_role key (PC only, in config.json) is what flips status.
Polled, not pushed — Task Scheduler: at logon + daily 02:00 with WakeToRun. No webhook, no listener, nothing exposed
03 · Worker — Windows PC, Python venv
run_auto.bat → pythonw library_tool.py --auto
pythonw = zero console flash. Per run: GET status=eq.pending with the service key, then for each row —
1 yt-dlp download, cookie chain --cookies-from-browser firefox → edge → chrome (Chromium cookies are DPAPI-locked on Windows; Firefox’s SQLite jar is readable, so Firefox must be logged in & closed).
2 Credit extraction from metadata — prefers a human uploader handle over Instagram’s numeric uploader_id; a Fix Credits pass can rewrite old entries from metadata alone, no re-download.
3 PATCH …?id=eq.{id} → status=done. Any failure: row untouched, retried next run. Every step timestamped into log.txt.
Storage — boto3 → Cloudflare R2 (S3 API)
put_object to the inspo bucket via the S3-compatible endpoint. Key = the reel’s shortcode (DaDpHOeuemh.mp4) — re-processing overwrites instead of duplicating, so the whole pipeline is idempotent.
list.json is read-modified-written alongside: an append-only array of {“video”, “handle”, “post”} — the entire “database” the public site ever sees. Zero egress fees; the R2 public domain is the CDN.
Front end never talks to Supabase or the PC — it only reads two static things: list.json and the MP4s
04 · Delivery — the wall, no framework
arboledaf.com/visualinfluence
Vanilla JS in one code block. fetch(list.json) → Fisher-Yates shuffle → one <video> tile per item in a CSS scroll-snap feed. An IntersectionObserver plays only the visible tile and lazy-assigns src from data-src, so 160+ videos don’t preload.
Autoplay policy handled per platform: muted autoplay playsinline attributes for iPadOS, sound upgrade attempted after, first-gesture unlock, silent retry on canplay — never more than one play prompt, ever. ended → auto-advance. Credit links back to the original post.