cursor/daily

Building with CursorCursor every single day

Admin
Manifesto

Code is easy now. Reality is still expensive.

I can vibe‑code anything in a day. The hard part is whether it survives real use. This log is me separating hype from signal and forcing one takeaway per build: what held up, what broke, what I’d change.

Inspiration↗
01Ship fast, then pressure‑test the weak parts.
02Personal tools over platforms — solve the now.
03No trophy builds — every project ends with a takeaway.
Today
 
 
57
Days
0
Streak
2026
Year
Building with
CursorCursorClaude CodeClaude CodeCodexCodexVercel
ManusManusConvexConvexSupabaseSupabaseOpencodeOpencodeTursoTurso
Latest takeaway
Thu, Feb 26
Daily takeaway begins Jan 16 — a lesson, failure, or mini blog note.
AMP Code Meetup conversations helped validate that small, tight agent loops produce faster feedback and better shipping momentum.
The vibe is optional. The takeaway is mandatory.
Sun
Mon
Tue
Wed
Thu
Fri
Sat
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec

Timeline

Jan 1

Built cursor/daily tracking website

8:53 PM
🤖Agent📄@Files🗂️@Codebase🔍@Web⚡Terminal💾Checkpoints🌐Browser📋Rules
3bd9515301bf5f975465c

Created a Next.js website to track daily Cursor usage with a 365-day dot-grid visualization, admin authentication, rate-limited login, and 24 Cursor features organized by category. Deployed to Vercel with Turso database. Live at https://cursor.bchewy.com

Jan 2

WhatToBuild AI startup idea generator

11:48 AM
🤖Agent💬Chat📄@Files⚡Terminal🎼Composer
ca1005d4449c10

Worked on WhatToBuild, a TanStack Start app that generates AI-powered startup ideas with a multi-step wizard flow and quick search feature. Live at https://ideas.bchewy.com/ PS: New to TanStack, wanted to try it out! Deployed on Cloudflare!

Jan 3

Bug fixes and test improvements

4:59 PM
🤖Agent📄@Files⚡Terminal
98984061a7e68cfd87bab

Fixed hydration error in dashboard greeting, various updates and test fixes.

Jan 4

Chrome Extension for LinkedIn Job Sync

6:30 PM
🤖Agent📄@Files⚡TerminalCursorClaude CodeCodex
1231060be27b28

Built Chrome extension for LinkedIn saved jobs sync with background service worker, MAIN world DOM injection, parallel tab processing, and blue-themed UI.

Chewremover AI Watermark Tool

7:09 PM
🤖Agent📄@Files⚡TerminalCursorClaude CodeCodex
02cb1260bff46d

Started Chewremover - a Next.js + FastAPI watermark removal app with AI-powered detection (corner, text, AI modes) and LaMa/OpenCV inpainting. Added CLAUDE.md, README, and package setup.

Cursor/Daily Timeline Enhancements

10:00 AM
🤖Agent📄@Files⚡TerminalCursor
27a9bb45802714

Enhanced the tracking website with horizontal scroll timeline, side-by-side multi-item layout, media thumbnails with lightbox, Jump to Today button, and SSR hydration fixes.

Jan 5

VibeCards AI Workshop App

4:40 PM
🤖Agent📄@Files📚@Docs⚡Terminal🔌MCPCursorCodex
51728c883eda0de11cb0c

Built a full-stack flashcard generator for AI workshop with Clerk auth, Supabase database, and OpenAI Structured Outputs. Features study mode, customization options, and fluid responsive design.

Jan 6

AI Pitch Coaching Platform

3:55 PM
🤖Agent🎼Composer📄@Files🗂️@Codebase⚡TerminalCursorCodexClaude Code
1c07cd04c2bdab

Built PitchHeat - a full-stack pitch coaching app where founders record 60-second pitches, get interrupted by an AI VC judge, and receive detailed scoring across 6 dimensions with shareable results. Features OpenAI Realtime API for voice, GPT-5.2 analysis, saliency-highlighted transcripts, and social-ready OG images.

Jan 7

Built ZipSwift iOS Puzzle Game

2:11 PM
🤖Agent🎼Composer📄@Files🗂️@Codebase⚡TerminalClaude CodeCursorCodex
a2db547e6cecac34817b117e1a1e7e9ece4c7c7d8e

Complete iOS puzzle game inspired by LinkedIn Zip. Features procedural Hamiltonian path level generation, 3 difficulty modes, synthesized audio with AVAudioEngine, haptic feedback, game history/statistics tracking, customizable settings with 8 accent colors, and polished SwiftUI interface.

Jan 8

YouTube Shorts Recipe Extractor MVP

2:09 PM
🤖Agent📄@Files🗂️@Codebase⚡Terminal🔌MCP🌐BrowserCursorClaude CodeCodex🎤Voice
1bce2c593735170c8c8c0

Full-stack app that extracts recipes from YouTube Shorts cooking videos using AI. FastAPI backend with OpenAI GPT-4o for recipe extraction, DALL-E 3 for food images, rate limiting, and 154+ tests. React frontend with auto-scrolling carousel, recipe history in localStorage, embedded video player. Docker deployment setup with nginx.

Jan 9

Zip Swift Mobile Game Showcase

12:12 AM
🤖Agent🎼Composer📄@Files🗂️@Codebase⚡Terminal🧠UltrathinkClaude CodeCursor👥Multi-Agent☁️Background Agent🔌MCP
a40e1dc7e845cca9757127891d3e3ef93ab3311069

Polished iOS puzzle game with statistics tracking, daily challenges, achievements, Game Center integration, widget support, and accessibility improvements. Full feature-complete showcase.

Jan 10

ROOMTHINK AR Room Scanner MVP

3:12 PM
🤖Agent📄@Files🗂️@Codebase⚡Terminal🔌MCPClaude CodeCursorCodex☁️Background Agent{ }@Code🌿@Git🧠Ultrathink🎼Composer
76e3f9c9680f12f684be42192ace

Built at the Gemini 3 Singapore Hackathon! An AI-powered room measurement app using Gemini Live API for real-time scanning. Implemented AR movement controls, 3D canvas with loading spinners, and interactive furniture scaling. Users can scan rooms, get dimension estimates, and generate interior design makeovers.

Hackathon organizers from @65labslah and @thorwebdev at Developer Space @ Google Singapore - not our team! Event: https://luma.com/gemini3sgp

Jan 11

Telegram Bot OpenAI Migration & New Commands

6:53 AM
🤖Agent📄@Files🗂️@Codebase📚@Docs⚡TerminalCursorCodex
220e7050ad7e357421a9b40a5b959030623

Complete rewrite of chewbot-prod, migrating from Gemini to OpenAI Responses API with gpt-5-mini. Added /tldr and /eli5 commands, token cost tracking, loading UX with in-place message editing, and auto-update on channel post edits.

Jan 12

Bullet Heaven roguelite game

2:25 PM
🤖Agent👥Multi-Agent📄@Files🗂️@Codebase⚡Terminal🔌MCPCodexCursor
47cf0e27c108fb1f6ad3a

Built a complete Vampire Survivors-style auto-attacking survival roguelite in Godot. Features 6 weapons, 5 enemy types, boss fights, wave scaling, meta-progression shop, save system, HUD, sound effects, and music.

Jan 13

Oneprompt Arcade

3:14 PM
🤖Agent🎼ComposerManus
bcd463c6ae4a1223e3e30d6eda6d56f1381

AI-powered arcade game generator built with Manus - describe a game in one prompt and it gets generated

Jan 14

TTS Throwdown: Blind Voice Benchmark

3:44 PM
🤖Agent🔍@Web📚@Docs⚡Terminal
8d4e5e65d177082959717

Built a blind A/B benchmark for text-to-speech providers featuring ElevenLabs, OpenAI, and Inworld. Users load prompts, listen to two voices without knowing the provider, vote for the winner, and a live leaderboard (powered by Turso) tracks results. Supports MP3 + WAV formats. Check it out at bchewy/tts-benchmark.

Jan 15

Built DoodleDo iOS doodle journal MVP

3:47 PM
🤖Agent📄@Files🗂️@Codebase⚡Terminal
33eaa66740e176

Created an iOS doodle journal app with PencilKit canvas, 2-column home grid with empty state, data model with in-memory store, and core tests via Swift package.

Jan 16

Manifesto hero + takeaway flow

2:19 AM
🤖Agent📄@Files🗂️@Codebase⚡Terminal
4a533dd

Redesigned the manifesto hero with a compact right rail and tool marquee, and wired per-item takeaways into the admin/editor flow starting Jan 16. Refined timeline/dot-grid behavior (horizontal scroll, jump-to-date, hidden scrollbar) to keep the daily log usable.

Takeaway
Vibe coding is cheap; forcing a takeaway keeps the work honest.

DoodleDo AI canvas refinements

2:19 AM
🤖Agent📄@Files🗂️@Codebase⚡Terminal
3e554c2

Refined the AI controls in DoodleDo iOS doodle journal - improved the canvas interaction for Draw/AI mode toggle, Full Canvas vs Lasso selection, and style presets like Cute.

Jan 17

Hack & Roll Hackathon Project 1st half!

4:08 PM
Claude CodeCodexCursorVercelOpencode

A simple cute project on hack n roll, on tauri!

Takeaway
Hackathons are about making weird ideas real. Turning keystrokes into music sounds absurd until you see chibi agents dancing to your debugging session. Ship the ridiculous - it's more memorable than another CRUD app.
Jan 18

Strudel VSCode Extension Visualizer

6:54 AM
🤖Agent📄@Files⚡Terminal
fdb8ac16515590

Built a visualizer and session recorder for the Strudel live coding music VSCode extension at Hack&Roll 2026.

Takeaway
Strudel is fun for live coding music, but next time exploring more serious applications like 2D-to-3D generation with Tencent Hunyuan model.
Jan 19

Personal commit tracker with email nudges

4:06 PM
🤖Agent🗂️@Codebase⚡Terminal🔌MCP
619ab9b4211066e4af2af

Built github-reminder - a tool that checks if you've pushed any commits today and sends motivational (or guilt-trip) emails if you haven't. Features password auth with HMAC tokens, GitHub API integration, and Vercel cron for automatic 6-hour checks.

Takeaway
Build tools that hold you accountable. Motivation fades but automated guilt-trips are forever. When you want to build a habit, automate the nagging so you can't ignore it.
Jan 20

Telegram Bulk Messenger Tool

4:06 PM
🤖Agent⚡Terminal📄@Files
8b1b44ca3f918644d84d1

Built a tool for sending bulk messages via Telegram.

Takeaway
Always build safety features first. Rate limiting, dry-run mode, and preview screens are not optional - they prevent catastrophic mistakes when working with bulk operations.
Jan 21

New features and updates

4:06 PM
🤖Agent📄@Files⚡Terminal
c9715d0

Added new stuff and improvements to justgritit.

Takeaway
Communities need a vibe, not just a landing page. Visual identity and playful elements like 3D worlds make people want to stay and explore.
Jan 22

PDF Slicer for Large Documents

3:52 PM
🤖Agent🎼Composer📄@Files⚡Terminal
1e89c12d923d303879802

Built a simple PDF slicer tool because uploading large PDFs to ChatGPT or Anthropic just does not work well - slicing is required for better processing.

Takeaway
When AI tools have limitations, build workarounds. Sometimes the best solution is to preprocess your data to fit the tool rather than fight against its constraints.
Jan 23

3D Chess Game with Three.js

7:31 AM
☁️Background Agent🤖Agent🗂️@Codebase
ee6fdf22daccf6aa8ba069b2f4e75d6d2ed9202cfede2637ff4bb3ac0a21aea876b073

Built a complete 3D chess game using Three.js and the background agent. Features include legal move validation, castling, pawn promotion, check indication, and game over detection.

Takeaway
Tested autoclaude for autonomous coding - it handled the 3D chess implementation but still lacks the polish of more mature loops like Ralph. Autonomous agents are promising for complex rule systems, but orchestration and context management still make or break the experience.
Jan 24

ZipSwift Bug Fixes

3:35 PM
Cursor🤖Agent
042d2f90361f3e6aaf665

Fixed the game state machine in ZipSwift - the win condition was not triggering properly when reaching the final node. Also fixed an edge case where players could keep drawing after completing the puzzle (but backtracking still works). Optimized path validation and fixed the level generator so new puzzles load without breaking.

Takeaway
Sometimes the hardest bugs are the ones hiding in plain sight. The win condition looked correct - but it was checking state at the wrong moment in the game loop. Lesson: when something "should" work but doesn't, question your assumptions about when code runs, not just what it does.
Jan 25

Semantic Sift - Telegram Archive Search

3:35 PM
🤖Agent📄@Files⚡Terminal🗂️@Codebase
02580ee9d9c92f4889fd4

Built a semantic search tool for Telegram chat archives. Instead of searching for exact words, it understands what you mean - search for "minecraft" and it finds conversations about game development and world-building, even if nobody said "minecraft." Makes years of chat history actually useful again.

Takeaway
Traditional search finds words. Semantic search finds meaning. The difference is like asking a librarian for "books with the word adventure" vs "books about adventure" - one matches text, the other understands intent. This changes how we can interact with our own data.
Jan 26

ZipSwift v1.1 - Sendable conformance & App Store launch

11:01 AM
🤖Agent📄@Files{ }@Code⚡Terminal
d73af99

Added Sendable conformance to GridPoint, LevelDefinition, GridSize, and Difficulty structs for Swift concurrency safety. Updated initializers and methods to be nonisolated where needed. Bumped marketing version to 1.1. ZipSwift is now live on the App Store!

Takeaway
  • Making structs Sendable-conformant is straightforward when they only contain value types — mostly just adding the protocol conformance and marking initializers as nonisolated
  • Cursor's agent mode handled the repetitive boilerplate of updating multiple model files with the same concurrency pattern efficiently
  • Shipping to the App Store with AI-assisted development: the code review and concurrency fixes that would normally take careful manual auditing were streamlined through agent + @code context
Jan 27

Inbox Atlas - Gmail Visualization

3:37 PM
🤖Agent📄@Files🗂️@Codebase⚡Terminal
68df49b203e120a36d38c

Built Inbox Atlas, a Next.js app that visualizes your Gmail inbox as an interactive constellation graph using D3. Connects via Google OAuth, maps email senders as nodes with relationship edges, and includes email censoring for privacy. Enhanced styling, added TypeScript definitions for D3, and refined the graph rendering logic.

Takeaway
  • Visualizing data you already have reveals patterns you never noticed. Turning a flat inbox into a constellation graph immediately surfaces your communication clusters and key contacts.
  • D3 force-directed graphs need careful tuning - link distance, charge strength, and collision radius all interact in non-obvious ways. Start with defaults, then adjust one parameter at a time.
  • Privacy-first features like email censoring should be built in from day one, not bolted on later. It changes how confidently you can demo and share your work.
Jan 28

Added walls and pause to ZipSwift

4:59 AM
🤖Agent📄@Files{ }@Code⚡Terminal
1b118d4

Added wall mechanics that block movement between adjacent cells, forcing alternate paths. Integrated wall generation into the level generator with difficulty-based counts. Also implemented pause/resume with accumulated time tracking. Bumped version to 1.2.

Takeaway
  • Walls add strategic depth by constraining valid paths — harder difficulties use more walls
  • Pause/resume requires tracking accumulated pause duration separately from elapsed time
  • Wall placement needs to be validated against the solution path to ensure puzzles remain solvable
Jan 29

ChatGPT Page Sender Chrome Extension

2:16 PM
🤖Agent📄@Files{ }@Code⚡Terminal
a81da248dc02bb3843218dd69fc71be7cc7fa67b2d9617732631fbf495a06833871177

Built a Chrome extension that extracts page content (visible text or full HTML) and sends it to ChatGPT with a custom or preset prompt. Features include prompt persistence, preset prompts, content mode toggle, and ProseMirror text injection.

Takeaway
  • Chrome MV3 extensions have strict content script sandboxing — session storage and chrome.action APIs behave differently than expected
  • ProseMirror editors (like ChatGPTs) need special injection handling via dispatch transactions rather than simple DOM manipulation
  • Building end-to-end in agent mode from scaffold to polish was smooth — breaking it into focused commits (scaffold, popup, injector, extraction, wiring, polish, fixes) kept each step manageable
Jan 30

Testing Kimi K2.5 on Mario Kart game

8:20 AM
🤖Agent💬Chat
4846113

Tested Kimi K2.5 model on a Three.js Mario Kart game project. It worked pretty well overall for game development tasks.

Takeaway
  • Kimi K2.5 handles game dev code (Three.js Mario Kart) competently
  • Good general-purpose coding ability for interactive 3D projects
Jan 31

SG IC Photo Compliance Web App

3:49 PM
🤖Agent📄@Files📚@Docs🔍@Web⚡Terminal
985024fd6680aa873cc07a0a7bc5

Built a Next.js web app to create Singapore IC/passport-compliant photos. Supports camera capture or file upload, guided cropping at 400x514px, and optional AI enhancement via Gemini 3 Pro Image for white background, lighting correction, and red-eye removal.

Takeaway
  • Gemini 3 Pro Image works well for photo enhancement tasks like background removal and lighting correction
  • Guided cropping with fixed aspect ratios gives users a much better experience than free-form cropping
  • HEIC/HEIF support is important for mobile users uploading from iPhone

Manus Hackathon

3:49 PM
Manus

Attended the Manus hackathon! Didn't win but had a great time exploring Manus and what it can do. The hack is still a work in progress. Made some new friends along the way!

Takeaway
  • Hackathons aren't just about winning — the real value is in exploring new tools and meeting people who share your curiosity
  • Manus is worth experimenting with; hands-on time at a hackathon is one of the best ways to understand a new tool's strengths and limits
  • Unfinished hacks are still valuable — the learning carries forward even if the project isn't polished yet
Feb 1

Contrarian Compass - Manus Hackathon

7:06 AM
🤖Agent📄@Files🗂️@Codebase⚡Terminal🔌MCP
47047ad15f883209ad48d1beb1039a29fd61e8036b95095c8a3cf670c802dde2825a8d6de0a02e4d0e92e78a0cd59bebc35c893df

Built Contrarian Compass, an AI-powered podcast analysis tool for the Manus hackathon. Analyzes transcripts to extract controversial claims, discovers people with opposing viewpoints via Exa and Perplexity, and generates personalized outreach angles for guest booking. Features interactive debate maps, sentiment analysis, and controversy scoring.

Takeaway
  • Multi-API orchestration (OpenAI + Perplexity + Exa) enables powerful claim extraction and counterpoint discovery pipelines
  • Force-directed graphs are effective for visualizing debate relationships between claims, people, and evidence
  • tRPC + React Query provides excellent DX for full-stack TypeScript apps with type-safe API calls
  • Hackathon velocity: agent mode + checkpointing enabled rapid iteration through UI redesigns and feature additions
Feb 2

Telebot dashboard migration to Next.js 15

7:54 AM
🤖Agent📄@Files🗂️@Codebase⚡Terminal📚@Docs
6824c0a7c974df

Telebot is a personal Telegram bot with a web dashboard for monitoring messages and managing bot state. Migrated the stack from Express + Vite React SPA to Next.js 15 App Router with TypeScript. Added real-time WebSocket updates, bcrypt password hashing, and new embedding utilities. Updated CLAUDE.md with full API/schema docs and fixed Traefik labels for Docker deployment.

Takeaway
  • Custom server.ts in Next.js 15 lets you attach WebSocket to the same HTTP server
  • CLAUDE.md with API tables and DB schemas makes agent collaboration much smoother
  • TypeScript migration catches API shape errors at build time
Feb 3

Ideagraph PDF knowledge graph app

3:09 PM
🤖Agent📄@Files🎼Composer
22a1c1a

Building a document visualization tool with PDF upload and interactive graph canvas for exploring document relationships.

Takeaway
  • PDF upload with drag-drop UX and 50MB limit
  • Graph canvas for visualizing document connections
  • Clean split-panel layout with documents sidebar
Feb 4

IdeaGraph PDF viewer, deployment, and promo video

3:33 PM
🤖Agent🎼Composer📄@Files🗂️@Codebase⚡Terminal📚@Docs
80d0eeee78432016337a08399e0f14454da

Built a PDF evidence viewer with highlight locators using react-pdf, added Docker and Traefik config for VPS deployment, added a mobile blocker overlay, and set up Remotion for promo video generation.

Takeaway
  • react-pdf with bounding-box highlights enables linking extracted evidence back to exact PDF locations
  • Multi-stage Dockerfile with standalone Next.js output keeps images lean for VPS deployment
  • Traefik labels in docker-compose simplify HTTPS routing without manual nginx config
  • Remotion compositions let you script promo videos as React components — great for product demos
Feb 5

Ideagraph open-source release prep

3:45 PM
🤖Agent📄@Files🎼Composer⚡Terminal
a4835432ae7d89000003c

Prepared ideagraph for open-source release with onboarding tour, UI polish, security hardening, and documentation.

Takeaway
  • Added onboarding tour for new users
  • Security hardening before going public
  • Root README for open-source discoverability
  • UI polish pass for release readiness
Feb 6

Ideagraph extraction quality and graph UX

12:35 PM
🤖Agent📄@Files🎼Composer⚡Terminal
1bc98be

Enhanced idea extraction with confidence threshold filtering and evidence count validation. Added document summaries to graph nodes, improved node entrance animations, and refactored layout calculations for group nodes.

Takeaway
  • Confidence threshold filtering improves graph signal-to-noise ratio by dropping low-quality extractions
  • Evidence count validation ensures only well-supported ideas survive into the graph
  • Document summaries on group nodes give context without opening the PDF
  • Node entrance animations make the graph feel responsive when new ideas land
Feb 7

Telebot searchable chat filter

6:00 AM
🤖Agent📄@Files⚡Terminal
197b785

Added a searchable combobox to replace the basic chat selector in the Telegram bot dashboard, making it easier to filter and find specific chats.

Takeaway
  • Combobox pattern with search is much better UX than a plain dropdown when dealing with many Telegram chats
  • Small focused PRs like a single component swap are quick wins for usability

IdeaGraph document folders and graph layout

10:00 AM
🤖Agent📄@Files🗂️@Codebase⚡Terminal
57c3fd3

Major update to IdeaGraph: added document folders with CRUD and bulk-move, overhauled DocumentList with search/filtering/status grouping, improved upload UX with progress bar, refactored graph layout from single horizontal strip to multi-row grid, added performance optimizations (memoization, conditional MiniMap, onlyRenderVisibleElements), and fixed multiple UI bugs.

Takeaway
  • Multi-row grid layout for React Flow scales much better than a single horizontal strip for many documents
  • Memoizing nodes/edges and using onlyRenderVisibleElements is essential for large graph performance
  • Key-based remount pattern cleanly fixes stale state in modals when switching between items

VMBoard internal status dashboard

3:30 PM
🤖Agent📄@Files🗂️@Codebase⚡Terminal
c1db2ef4e91804

Built a Next.js 15 internal dashboard for monitoring Firecracker VMs and host Docker containers. Tracks two VMs (brianbot, justgritit) and 12 host containers with live status, log viewing, and health checks. Deployed via Docker with Traefik routing and Cloudflare DNS challenge for TLS on Tailscale.

Takeaway
  • Dockerode library works well for monitoring container status from within a sibling container via Docker socket binding
  • Firecracker VM monitoring over SSH adds useful visibility into microVM health
  • Tailscale-only deployment with Cloudflare DNS challenge keeps internal tools off the public internet
Feb 8

Firecracker VM control plane for OpenClaw

8:28 PM
🤖Agent⚡Terminal📄@Files🗂️@Codebase
1a13f900e98435191b892c43e64ba92908d

Built a minimal control plane for running OpenClaw instances inside Firecracker microVMs with full isolation. Includes vm-setup for provisioning (rootfs, cloud-init, networking, Docker), vm-ctl for lifecycle management (start/stop/restart/destroy/logs/ssh), and an agent skill following the Open Agent Skills standard. Each VM gets its own filesystem, network, and process tree with ~125ms boot time and ~5MB overhead.

Takeaway
  • Firecracker microVMs provide strong isolation with minimal overhead (~5MB per VM, ~125ms boot) compared to Docker containers
  • cloud-init with seed images is a clean way to bootstrap VM guests without baking custom images
  • Open Agent Skills standard provides a structured way to expose VM control as an agent-callable skill
  • Plain bash + systemd is sufficient for a lightweight VM control plane without needing Kubernetes
Feb 9

PhishGuard phishing awareness training game

1:35 PM
🤖Agent📄@Files⚡Terminal📋Rules
e3e6ab1

Built PhishGuard, a Next.js 16 phishing awareness training game where players analyze realistic email scenarios in a simulated email client. Features 12+ scenarios across 4 categories, scoring/health system, email verification tools, and an LLM-powered endless mode using OpenAI Responses API. Includes a Remotion-generated demo video.

Takeaway
  • Next.js 16 SPA mode works well for game-like interactive apps with complex client state
  • Remotion makes it easy to create polished demo videos entirely in React
  • Zustand + localStorage provides a clean persistence pattern for game progress without a backend database
Feb 10

GitLab support for cursor tracker

2:22 PM
🤖Agent📄@Files⚡Terminal
0d78b931394f2a

Added GitLab support so the daily Cursor usage tracker can show commit links for GitLab repos as well as GitHub.

Takeaway
  • Extended commit link logic to detect GitLab repo URLs and use GitLab commit URL format
  • Tracker now supports both GitHub and GitLab repos in the same entry format
  • Single source of truth for repo/commit display keeps UI consistent across hosts
Feb 11

Fred Again.. Apple Music Roulette

3:11 PM
🤖Agent🔍@Web⚡Terminal📄@Files
6f1eba31646c423c79901

Built a Fred again.. roulette web app that fetches songs from the iTunes Search API, spins a CSS roulette wheel to pick a random track, and plays 30-second previews with album artwork. Includes a Remotion-powered showcase video.

Takeaway
  • Vanilla JS + CSS can ship a polished roulette UX without any framework overhead
  • iTunes Search API is free and returns up to 200 results with 30s preview URLs
  • Remotion showcase video adds a nice demo artifact alongside the live app
Feb 12

OpenAI Fundocs - Plain Language Docs

3:39 PM
🤖Agent📄@Files📚@Docs🔍@Web⚡Terminal
0693b49

Built openai-fundocs, a rewrite of the OpenAI platform documentation aimed at non-technical readers. Uses workshop metaphors, plain language explanations, and custom MDX components (Callout, PromptSandwich, ToolPicker, AgentBriefBuilder) to make AI concepts accessible. Covers core concepts, agents, and tools across 30+ pages with Next.js 16, Tailwind v4, and next-mdx-remote.

Takeaway
  • Plain language rewrites of technical docs require thinking about "what happens next" and "what can go wrong" rather than just API parameters
  • MDX with custom components (PromptSandwich, ToolPicker) bridges the gap between docs and interactive learning
  • Workshop metaphors (Writing Desk, Form Filler, Errand Runner) make abstract AI concepts tangible for non-technical stakeholders
  • Next.js 16 + Tailwind v4 + next-mdx-remote is a solid stack for content-heavy docs sites
Feb 13

Background Agent ships 82-commit Telegram bot PR

3:14 PM
☁️Background Agent🤖Agent🗂️@Codebase
a6600cf547c7b9ab6b03472fd82465e36d5edd4da29c01237ebc58df9c01a1e9889149

Cursor's Background Agent ran a single autonomous session and produced 82 commits with 4,608 lines of code for Chewbot, a Telegram group bot. The agent built the entire feature set end-to-end: dashboard auth, anon aliases, hype meter, per-user AI rate limiting, weekly digests, role-based access (admin/mod/member), reaction-based remixes, model management with setmodel/audit history, state persistence, and then went deep on security - building out command access audit logs, deny-rate telemetry, security posture heuristics, configurable alert thresholds, and filterable audit/denied endpoints across both bot commands and dashboard API. All 82 commits co-authored by [email protected]. Zero intervention.

Takeaway
  • Cursor Background Agents can run autonomously for extended sessions producing 82 commits and 4.6k lines with zero babysitting
  • The agent naturally progressed from core features to security hardening without being told - it built audit logs, deny-rate telemetry, and security alerts on its own
  • Fire-and-forget workflow: one prompt produced a complete PR with bot logic, dashboard API, RBAC, and full security infrastructure
  • Co-authored-by [email protected] pattern makes the collaboration visible in git history
Feb 14

11h42m Background Agent: 250 commits, 16k lines

11:41 AM
☁️Background Agent🤖Agent🗂️@Codebase
a6600cf9889149ff4e9a72bc26a43039ff361ee504bf742b6e9339d06d864d0c2cf6d2

Cursor's Background Agent ran for 11 hours and 42 minutes in a single autonomous session on Chewbot, producing 250 commits and 16,355 lines of code. Starting from yesterday's 82-commit foundation, the agent kept going - first building out features (model management, security alerts, deny-rate telemetry, configurable thresholds), then autonomously shifting into a massive refactoring phase: extracting helpers, creating dataclass query parsers, centralizing scope parsing, modularizing the dashboard and bot into clean service layers. It essentially wrote the app, then refactored it into production-quality architecture. All commits co-authored by [email protected].

Takeaway
  • Background Agents can run for 11h42m producing 250 commits and 16k lines autonomously - this is overnight coding on autopilot
  • The agent self-directed through phases: features → security hardening → refactoring → modularization, without being prompted for each phase
  • It wrote the code AND cleaned it up - extracting helpers, creating dataclasses, centralizing parsers - things you'd normally do in a separate PR
  • 250 commits from a single prompt makes background agents viable for shipping entire features end-to-end while you sleep
Feb 15

Fix Fireclaw SSH instance ID handling

4:41 PM
🤖Agent📄@Files⚡Terminal
d0a3bc0

Fixed SSH wait checks in Fireclaw (Firecracker microVM control plane for OpenClaw) to use explicit instance IDs instead of implicit references, preventing misrouted SSH connections when managing multiple VMs.

Takeaway
  • Explicit instance ID passing prevents subtle bugs when orchestrating multiple Firecracker VMs
  • Small targeted fixes matter in infra code where implicit state can cause hard-to-debug misrouting
Feb 16

Mini Codex — Tauri desktop GUI for Codex CLI

3:26 PM
🤖Agent
b55e2bd

Built a Tauri 2 desktop app with React 19 and Rust that wraps the OpenAI Codex CLI. Features a chat interface, model selection (GPT-5.3, o3, o4-mini), sandbox mode toggles, YOLO mode, HTML file previewer, and real-time streaming of Codex output via Tauri events.

Takeaway
  • Tauri 2 + React 19 is a solid combo for quickly shipping native desktop apps with web UIs
  • Rust backend handles spawning and managing CLI subprocesses with real-time event streaming to the frontend
  • Wrapping an existing CLI tool with a GUI makes it far more accessible without rewriting core logic
Feb 17

IssueRadar search and UX polish

1:40 PM
🤖Agent💬Chat📄@Files🌿@Git⚡Terminal
e2b5bc148951cb

Built IssueRadar, an AI-powered GitHub issue and PR search app, then polished both backend APIs and frontend UX. Added OAuth support, repo validation/search improvements, loading states, presets, and a stronger README for onboarding.

Takeaway
  • Relevance scoring plus clear "why this matches" context makes issue triage much faster
  • Backend caching/logging and frontend loading states together improve perceived speed and trust
  • Tight docs + UI polish in the same pass reduces onboarding friction for new users
Feb 18

Fireclaw provision, tests, and docs

12:37 PM
🤖Agent⚡Terminal📄@Files🗂️@Codebase
8528164d61d0905b13d58

Added a provision command with host-port flag and hardened allocation helpers. Wrote unit tests for vm-common.sh. Rewrote README, llms.txt, and added AGENTS.md for agent-friendly context.

Takeaway
  • Provision command rounds out the CLI for full VM lifecycle management
  • Unit tests on shell helpers catch edge cases early
  • AGENTS.md + llms.txt make the repo agent-ready for future sessions
Feb 19

Redesigned Fundocs with dark glass aesthetic

3:42 PM
🤖Agent📄@Files🗂️@Codebase
3d849dc

Rebuilt the OpenAI Fundocs site with a pure black background, blue accent palette, frosted glass card surfaces, and Syne + Manrope typography. Covers Core Concepts, Agents, and Tools in a non-technical-first approach.

Takeaway
  • Dark theme + frosted glass cards create a premium feel for docs sites
  • Non-technical-first framing makes API docs more accessible to broader teams
  • Syne + Manrope pairing works well for bold headers with readable body text
Feb 20

Luma Hackathon Review Dashboard MVP

2:51 PM
🤖Agent📄@Files⚡Terminal🗂️@Codebase
19280d79e3236d

Built a hackathon review console for managing OpenAI Codex Hackathon Singapore applications. Handles 525 pending applicants across 438 teams with bulk approve/decline workflows, team detail panels, name variant detection, and filtering by event, status, and registration type.

Takeaway
  • Shipped a full admin review dashboard from scratch in a single session using Next.js
  • Luma API integration for pulling hackathon registrations and managing approvals at scale
  • Name variant detection helps catch duplicate team registrations
  • Dark-themed data-heavy UI with stats cards and tabbed team/people views works well for review workflows
Feb 21

Luma Review Dashboard Enhancements

2:49 PM
🤖Agent📄@Files⚡Terminal🗂️@Codebase
6d480f8

Iterated on the hackathon review console — added persistent filter/panel settings, draft decisions with export tools, improved member detail interactions, and normalized social profile URLs.

Takeaway
  • Persisting filter state across sessions is a small touch that saves reviewers a lot of friction
  • Normalizing social URLs (handling missing leading slashes) is the kind of edge case that bites you in prod
  • Draft decisions + export workflow lets reviewers batch work without committing immediately
Feb 22

Cloudflare R2 media offload rollout

1:50 PM
🤖Agent📄@Files🗂️@Codebase🌿@Git⚡Terminal🔍@Web
494408f949eaf9

Migrated cursorbrian media delivery from Vercel-served /public paths to Cloudflare R2 via media.bchewy.com. Added runtime URL resolution for existing media paths, uploaded media to R2, attached custom domain, updated Vercel env, and verified production now serves cached media from Cloudflare.

Takeaway
Decoupling media delivery from app hosting is an immediate cost win. If your data already stores relative media paths, a runtime URL resolver gives a near-zero-friction CDN migration.
Feb 23

Redacted Private Project Work

11:35 AM
🤖Agent📄@Files🌿@Git⚡Terminal

Worked on a redacted private project today.

Takeaway
Making preview seeding opt-in and restoring ingestion unmount behavior cut stale-state regressions while keeping strict typechecks green during fast setup-flow changes.
Feb 24

Modal training + inference app

3:03 PM
🤖Agent⚡Terminal📄@Files📁@Folders
92d07c2

Built an end-to-end Modal training pipeline and inference web app for the Telegram classifier dataset, connecting training outputs directly into a usable prediction surface.

Takeaway
Combining pipeline + app in one iteration tightened the feedback loop and exposed integration issues earlier than doing them separately.
Feb 25

Realtime Reflective Mirror MVP

3:13 PM
🤖Agent⚡Terminal📄@Files🌿@Git
3b91a93

Built the first MVP for a realtime reflective mirror companion and wired the core interaction flow end-to-end.

Takeaway
Keeping the first pass tight around one user loop made integration and debugging significantly faster.
Feb 26

AMP Code Meetup Notes

3:22 PM
💬Chat🤖Agent📚@Docs🔍@Web⚡Terminal
3b91a93

Attended the AMP Code Meetup and spent time around practical agent workflows and iteration speed. Captured ideas to translate meetup learnings into tighter build loops for upcoming realtime work.

Takeaway
AMP Code Meetup conversations helped validate that small, tight agent loops produce faster feedback and better shipping momentum.
Scroll to explore