Thinker

Introduction

Thinker is a Progressive Web App for intentional microlearning. Instead of an infinite scroll, it presents one card at a time — curated ideas, ingested news, scripture, free learning sites, Project Gutenberg books, and short brain games — then pushes you toward the real source when you want depth.

Live: https://thinker.360web.cloud
Source: mitchelldawkinsjr/thinker

For the product narrative, see the companion post: Thinker: Replace Doomscrolling with Microlearning.

Status: Active — Deployed on 360 Web Solutions Cloud with weekly content ingest and optional Ask (OpenAI preferred, Ollama fallback).

Problem & Solution

The Problem

Doomscrolling wins by default — feeds optimize for time-on-app, not residue of learning
Read-later piles never shrink — bookmarks and RSS become another inbox
Generic AI chat invents links — answers without a trustworthy outbound shelf
Accounts and sync are overkill for a personal daily practice

The Solution

One-card deck — mixed feed with weighted interleave across content kinds
Day-local memory — cursor and “seen” reset each calendar day; Kept/hidden survive
Catalog-grounded Ask — instant keyword path, then LLM refine restricted to known URLs
Static + ingest — content in-repo / JSON files; no per-user database
Self-hosted PWA — Docker + nginx on VPS, installable to the home screen

Architecture Overview

Layer	Implementation
Client	React 19, react-router, TypeScript, Vite
Install / offline	`vite-plugin-pwa` (service worker + webmanifest)
Content	`src/data/*` + `public/content/{news,scriptures,book-ideas}.json`
Edge	nginx SPA routing, healthz, aggressive `/assets/` cache
Ask	`/api/openai/` (server-side key) · `/api/ollama/` · instant `exploreInstant()`
User feeds	Node sidecar `/api/feed-proxy` for CORS-safe RSS
Deploy	`.github/workflows/deploy-vps.yml` → Docker Compose on VPS
Ingest	Weekly Actions: news + scriptures + book ideas → commit JSON

Mixed Feed

The feed is assembled client-side by buildMixedFeed():

Item kinds: idea · news · scripture · resource · book · game

Key behaviors:

Weighted interleave — ideas/news pulled more often without duplicates
Seeded shuffle — stable-enough day mix; Reshuffle bumps a key
Unseen rises / news expires — freshness without an engagement algorithm
Topic filter — URL ?topic= plus subscription affinities
Hide — permanent dismiss from the pool; Keep — bookmark for later

Thinker feed UI with idea card, Keep and Ask actions, and mix counts

Day Session & Device Memory

No auth. Persistence is localStorage with an intentional split:

Today-only day key cursor and seen versus overnight kept hidden and subscriptions

Key / concern	Lifetime	Role
`thinker-day-v1`	Today	Calendar day marker; roll clears session memory
`thinker-day-cursor-v1`	Today	Card index per topic for Continue CTA
Seen / rotation	Today	Avoid immediate re-shows in the mix
`thinker-kept`	Forever	Bookmarked card IDs
Hidden	Forever	Permanently removed from pool
Subscriptions	Forever	Custom sites / feeds / preferences

ensureFreshDay() runs on open: if the local calendar day changed, wipe seen + cursor; leave Kept and hidden alone.

Ask: Instant + LLM Refine

Ask pipeline from question through instant catalog match to OpenAI or Ollama refine

Instant path (`exploreInstant`)

Tokenize question + optional idea/topic context
Score topics, browseable resources, Gutenberg meta
Always return at least one catalog link (SEP fallback)
Target latency: under 5ms

LLM refine (same JSON contract)

{
  "answer": "max 50 words",
  "digDeeper": ["q1", "q2"],
  "links": [{"title": "t", "url": "https://...", "why": "short"}],
  "topics": ["topic-id"]
}

Rules enforced in the system prompt: use only catalog URLs, never invent links, no social-feed URLs, tight token budgets (max_tokens / num_predict ~220).

Provider	When	How
OpenAI `gpt-4o-mini`	Preferred when `OPENAI_API_KEY` set	nginx proxies to `api.openai.com`; key not in JS
Ollama	Fallback	`/api/ollama/` → host/VPS runtime
Instant only	Neither available	App still fully usable

Topics & Content Model

Fourteen built-in topics (AI agents, LLMs, RAG, AI frontend craft, NBA, WNBA, football film, sports biz, current events, history, politics, finance, mental models, building products) each with color/accent and curated idea pools.

An Idea carries hook → body/lesson → takeaway → example, plus source / sourceUrl / optional gutenbergId and readMinutes.

Weekly CI (ingest-news.yml) refreshes:

public/content/news.json
public/content/scriptures.json
public/content/book-ideas.json

Production Deploy

Item	Detail
URL	`https://thinker.360web.cloud`
DNS	CNAME `thinker` → `360web.cloud`
Runtime	Docker (`Dockerfile` multi-stage: Node build → nginx Alpine)
Network	Shared `360ws-network` with sibling 360web apps
Health	`GET /healthz` → `ok`
Secrets	`VPS_HOST`, `VPS_USER`, `VPS_SSH_KEY`; runtime `.env` for OpenAI/Ollama

cd /opt/thinker
docker compose -f docker-compose.prod.yml up -d --build

PWA assets (sw.js, manifest.webmanifest) are served with Cache-Control: no-cache; hashed Vite assets are immutable long-cache.

Key Features

Installable PWA — Add to Home Screen; offline shell
One-card microlearning deck — mixed content, not a timeline
Continue today — resume mid-feed on the same device
Kept shelf — durable bookmarks without an account
Resources + Gutenberg — curated outbound destinations
Ask — instant catalog paths + optional cheap LLM refine
Custom subscriptions — user sites/feeds via feed proxy
Self-hosted cost curve — static app + scheduled ingest

Thinker on mobile showing hero and floating idea cards

Local Development

git clone [email protected]:mitchelldawkinsjr/thinker.git
cd thinker
cp .env.example .env   # optional — Ask (OpenAI or Ollama)
npm install
npm run dev

Optional Ask env:

OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-4o-mini
# or
OLLAMA_URL=http://127.0.0.1:11434
VITE_OLLAMA_MODEL=phi3:mini

Without either, Feed / Topics / Resources / Kept still work; Ask uses the instant catalog path.

Who This Is For