Prabha · 2026-06-07 · All times IST
| Time | Category | Format |
|---|---|---|
| 09:00 | workflow | single |
| 10:30 | hermes | thread (5 parts) |
| 12:00 | stats | single |
| 14:00 | failures | thread (3 parts) |
| 15:00 | apps | single |
| 16:30 | apps | single |
| 18:00 | tips | single |
| 19:30 | interesting | single |
Copy each tweet at the scheduled time. Engage with replies within 30 min. Threads: post parts as replies, not separate tweets.
My design review workflow now: 1. Agent captures screenshots 2. HTML gallery generated automatically 3. I add commentary inline 4. Gallery pushed to GitHub Pages 5. Share the link. Done. No Figma comments. No email threads. No "check the latest version." One URL. One source of truth.
Why I built my own agent instead of using ChatGPT: A thread on persistent memory, custom skills, cron jobs, and running AI for under $5/month.
Persistent memory. ChatGPT forgets everything when the conversation ends. My agent remembers my project structure, preferences, and past failures. I said "push to TestFlight" today. It already knew my scheme, bundle ID, and API key path. No re-explaining.
Custom skills for MY workflow. Not "help me write code." Specific: "build WhatPay for TestFlight using my signing config." I have 30 skills. 5 get used daily. Skills are not Pokemon. You don't collect them. You weaponize the ones that remove friction.
Cron scheduler. Monday 9 AM: ITSM research runs automatically. I wake up to a finished digest. Not a reminder to do research. The research is already done. Daily tweets, weekly competitor scans, overnight research. All run without me touching anything.
Cost: ~$5/month. Claude API for heavy lifting. Ollama for daily driver. Self-hosted search and scraping. Data stays local. No training on my code. No retention of my financial data. Total control: 100%. Total subscription fees: $0.
My agent setup by the numbers: 30 skills (5 used daily) 6 cron jobs running 5 iOS apps in the store ~$5/month total cost Most-used: Todoist processing, iOS builds, ITSM research, design guardrails, HTML publishing. The system learns what I use. The rest gathers dust. That's fine.
Google OAuth broke on my agent setup. Every API call failed with 403 access_denied. Went through the entire OAuth flow twice. New credentials, new refresh token. Still 403. The error said "client not authorized." That was a lie.
Turned out: my app was in "Testing" mode in Google Cloud Console. In testing mode, only explicitly added test users can access it. My own email wasn't on the list. One checkbox. Add test user. Done. Months of it working fine, then suddenly not.
The error said "access_denied" and "client not authorized." It did NOT say "user not in test list." Google auth is the YAML of authentication: looks simple, hides complexity, and the error messages actively mislead you.
5 apps in the App Store. Every one started as "this is annoying, let me fix it." WhatPay - expense tracking without the phone number TimeLeft - countdown widget that looks good LeaveHack - PTO tracker (HR portal is unusable) Sol Vilaiyatu - Tamil word game for my parents ThatsMyFont - custom handwriting font None came from market research. All came from friction.
Shipped WhatPay v1.3.2 last night. Removed privacy/terms clutter from settings (8 rows to 4). Rewrote paywall copy from feature-list to outcome-based. Build pipeline: one command, 30 seconds, TestFlight notification before I finish my coffee.
Keep a .gitignore template for iOS projects. I committed build artifacts and DerivedData into a repo. 2 MB ballooned to 87 MB. Had to rewrite git history. Now I copy ~/.gitignore-ios into every new project on day one. Costs 10 seconds. Saves hours.
Most "AI automation" content is generic garbage. "Use ChatGPT for emails!" "Automate your calendar!" The real stuff nobody talks about: - Cron-scheduled research pipelines - Custom skills that encode YOUR workflow - One-command TestFlight uploads - Persistent memory that learns from failures That's actual automation.