EPISODE 03 · AI tools & workflows · Rewritten 2026.07

22 workflows,
running on my Mac right now.

This page used to be about methodology. It's now all real artifacts —
the workflows I run every single day, laid out one by one.

Start with the scheduled routines →
AGENT 101 SERIES · EP 03 · ALL RUNNING FOR REAL
§ 01 · Division of labor first

In this whole system, the human does four things

The 22 items below aren't 22 isolated tricks — they're one system where each piece feeds the next. Before reading any of them, see how the work is split between me and the AI: I spend less than half an hour a day on "operating" anything.

  • Tick boxes. Every morning I scan the Obsidian cockpit page and check the material worth keeping. Everything else gets ignored.
  • Paste links. When I hit something good during the day, I paste the link into a collection box. A nightly job files it automatically.
  • Record. I pick a script from the backlog, read it on camera, and upload the video to Google Drive.
  • Make the call. Publish or not, build or not, buy or not — business judgment always stays with the human.
§ 02 · Fully automatic · Scheduled routines

Six routines, running on a timetable all day

These six run as local scheduled tasks in Claude. If the Mac is on, they run themselves; if it's off, they catch up next boot. From 5:30 AM to late night, the entire input side — collecting and filing — is unattended.

  • ROUTINE · 01 · Daily 05:34 0 minutes of human time

    AI news archive: the briefing is written while I sleep

    Before I wake up, the last 24 hours of AI news has already become a categorized briefing, filed in my knowledge base.

    ROUTINEdaily-ai-hotspot-collect
    01PULLFetch the past 24h of curated AI news from a trends API
    02WRITEDraft the briefing · grouped by category · source link on every item
    03IDEASExtract 3–5 content-topic ideas · angle + audience + format
    04FILESave into the Obsidian dailies archive
    Claude Code Trends API Obsidian
  • ROUTINE · 02 · Daily 06:11 One screen over breakfast

    Morning cockpit: the whole day's information on one page

    Three news sources plus one inbox, assembled into a single Obsidian homepage every morning. All I do is tick boxes while eating breakfast.

    ROUTINEmorning-cockpit
    01FETCHThree sources · AI news / GitHub trending / non-AI headlines
    02DEDUPCheck against the ledger · yesterday's items never reappear
    03EMAILTriage overnight email into three buckets via the Gmail connector
    04BUILDAssemble the homepage · material to approve + today's to-dos, one screen
    Claude Code Gmail MCP Obsidian
  • ROUTINE · 03 · Daily afternoon +1 spare script per day

    Daily short-video script: the backlog always outruns my filming

    One 90-second talking-head script per day, in both English and Chinese. The Chinese isn't translated from the English — it's rewritten for how spoken Chinese actually sounds.

    ROUTINEshort-video-daily
    01TOPICPick from the AI-learning-methods topic pool
    02DRAFTWrite a 90-second word-for-word script in each language
    03CHECKOpening passes a hook check · Chinese passes an AI-flavor check
    04FILEInto the backlog · I choose at filming time
    Claude Code Custom skill Obsidian
  • ROUTINE · 04 · Mon / Thu Two 15-minute scripts per run

    Long-video scripts: topics come from formulas mined out of 101 hit videos

    Topic selection doesn't rely on inspiration. I first deconstructed all 101 videos from two top creators into a library of topic formulas — the routine drafts from those formulas.

    ROUTINElong-video-2x-week
    01FORMULAPick a topic from the formula bank · check the ledger for repeats
    02DRAFT2,000+ word English script · Chinese written separately, not translated
    03VERIFYEvery number in the script gets a source link
    04FILEInto the library · I filter at filming time
    Claude Code Topic formula bank Obsidian
  • ROUTINE · 05 · Nightly 21:37 One tick = one item filed

    Nightly ingest: what I ticked in the morning becomes knowledge cards at night

    The items I checked at breakfast and the links I pasted during the day get processed together every night: articles fetched in full, videos transcribed, everything filed as knowledge cards by domain.

    ROUTINEevening-ingest
    01READCollect ticked items from the homepage + links from the inbox page
    02FETCHArticles: extract full text · videos: transcribe with Whisper
    03CARDDetect the domain · generate a knowledge card from the template
    04LEDGERUpdate the ledger · the same item is never filed twice
    defuddle Whisper Obsidian
  • ROUTINE · 06 · Wed 03:00 450+ saved posts being digested

    Favorites harvest: turning "saved it once" into "can actually look it up"

    All the Xiaohongshu and Douyin posts I saved and never opened again get pulled weekly, filtered for actual teaching content, and transcribed into the knowledge base.

    ROUTINEweekly-favorites-harvest
    01SCRAPEPull both platforms' favorites lists via the logged-in browser
    02FILTERKeep teaching content only · pure entertainment is skipped
    03SCRIBETranscribe videos · transcribe image posts frame by frame
    04CARDFile knowledge cards · media files deleted after transcription
    Claude in Chrome yt-dlp Whisper
§ 03 · Manually triggered · Content pipeline

From one trigger phrase to an auto-edited video

The scheduled routines stock the raw material; this group turns it into finished videos. The human appears exactly twice: approving the topic, and reading the script on camera.

  • PIPELINE · 07 · Trigger: one phrase One run = several finished scripts

    Knowledge base → talking-head scripts

    I say one phrase — "run a batch" — and it benchmarks against reference channels, picks a topic, checks the ledger for repeats, then produces several word-for-word scripts from one topic: a 15-minute long script plus a 1-minute vertical short, written separately in English and Chinese. The hard rule for every script: I read it on camera as-is, so nothing unreadable-aloud is allowed — even numbers are written out as words.

    Topic bank Hook check AI-flavor check
  • PIPELINE · 08 · Trigger: new file in Drive Shoot, upload, done

    Post-shoot auto-editing

    I film the script and drop the video into a raw folder on Google Drive. A daily job detects the new file and takes over: it locates every paragraph by transcript timestamp and overlays illustrations and B-roll cards paragraph by paragraph. Scripts are written as "one paragraph = one visual beat" precisely so this step can be fully automatic.

    Google Drive Transcript timestamps Auto B-roll overlay
  • PIPELINE · 09 · Trigger: paste a link Competitor research without memory

    YouTube / Xiaohongshu / Douyin video → transcript

    See a good video, throw it the link: YouTube and Douyin videos get downloaded and transcribed with Whisper; Xiaohongshu image posts get transcribed including the text inside the images; everything is structured into a document and filed in Obsidian. This is exactly how those 101 videos from two top creators got fully deconstructed — which is where the topic formula bank above came from.

    yt-dlp Whisper Obsidian
  • PIPELINE · 10 · Trigger: name an industry No more guessing at topics

    Viral topic generator

    Give it an industry and it produces a batch of topics from validated formulas. Feed it "home renovation" and you get "the furniture that loses value fastest" and "the least impressive renovation style" — topics with built-in controversy. The formulas hold across industries.

    Topic formulas Any industry
§ 04 · Browser control · Website & leads

Not giving advice — finishing the job

What this group has in common: Claude doesn't tell me how to do things in a chat window. It drives the browser and does them. I only step in at the moments that genuinely require a human — typing a verification code, recording a verification video.

  • SESSION · 11 13 pages shipped

    Local-business website: written from zero, deployed

    Built the site for my Dallas web-services business: 13 pages of pure HTML, deployed straight to Cloudflare Pages. No framework, no build chain — a small business's website should be exactly this light.

    Pure HTML Cloudflare Pages
  • SESSION · 12 Zero email lost

    Moving DNS wholesale

    Migrated a domain from GoDaddy to Cloudflare entirely in the browser. Before the move it inventoried every Microsoft 365 email record into a checklist, then verified each one after — the scariest part of switching providers is your email silently dying.

    GoDaddy Cloudflare DNS M365 record inventory
  • SESSION · 13 The ticket into local search

    Registering a Google Business Profile

    The whole Google Business Profile flow, driven in the browser: picking categories, setting a service area, going the no-storefront route. At the step that genuinely requires a human — video verification — it parked the page exactly where I needed to take over, and continued after I finished.

    Google Business Profile Video verification
  • SESSION · 14 Brand terms fully indexed

    Five rounds of SEO + page-by-page index requests

    Sitemaps, structured data, long-tail keyword strategy — five documented iterations. After new pages ship, it requests indexing in Google Search Console page by page until it hits the daily quota.

    Search Console Structured data Long-tail strategy
  • SESSION · 15 Taste, but engineered

    Redoing the whole site's design with design skills

    I didn't like the first version of the design, so I installed two design skills and started over: one enforces taste red-lines (type scale, whitespace, banned clichés), the other maintains the design system. The output design spec is persisted as a document — every later page follows it.

    Design skills Design-system doc
  • SESSION · 16 Same playbook, reusable

    SEO / GEO for someone else's site

    Took the same process to a friend's website and ran it again: diagnose, fix, submit for indexing. The first time is exploration; the second time it's an SOP — which is exactly the line between "a workflow" and "a one-off."

    SEO GEO Repeatable SOP
§ 05 · Decisions & odd jobs

The rest don't form a system — but they all genuinely save time

Not every workflow deserves an architecture diagram. This group is loose parts — each solves one small thing, but together they're what "AI in daily life" actually looks like.

  • ODD JOB · 17 · The star of this site's Issue #1

    Email → labels + a to-do list

    Claude Pro plus the Gmail connector reads my inbox once a day: every email gets labeled Action Required / FYI / Newsletter / deletable, then it outputs "the three things that must happen today" and drafts the replies that need answering. This was the first workflow I ever built, and the case study in Issue #1 of this site — the lowest barrier to entry of all 22, and the best first workflow for you too.

  • ODD JOB · 18 · Meetings

    AI meeting notes: bring only your brain into the room

    Record the whole meeting, transcribe it afterwards, and have Claude structure it into minutes: what was decided, who owns it, when it's due. In the meeting I only listen and judge — the pen never moves.

  • ODD JOB · 19 · Business decisions

    A business-diagnosis skill: taking "should I do this" apart

    Questions like "should I sell a course" or "what should I charge" don't get settled by feel. A custom diagnosis skill walks them through a framework: audience, pricing, delivery, opportunity cost. I don't always follow its verdict — but it forces me to finish the math.

  • ODD JOB · 20 · Paperwork review

    Rental application screening

    When applications come in for my rental, it reviews each file: income verification, red flags annotated, follow-up questions listed. The human reads the summary page and makes the final call.

  • ODD JOB · 21 · File chores

    Ebook conversion, sent to Kindle

    One sentence: convert this book to a Kindle-friendly format and send it to my Kindle email. What used to mean tutorials and installing software is now a single message.

  • ODD JOB · 22 · Self-archiving

    Session log, structured and emailed to myself

    When an important working session ends, it writes the whole process up as a structured email to my inbox. The decision process itself is an asset — worth archiving.

§ 06 · Want to copy this system?

Don't start with 22. Start with 1.

This system wasn't built in a weekend — it grew one workflow at a time, and anything that didn't run smoothly got deleted. Follow this order. Don't get greedy.

  • Week 1 · Plug in your first one

    Start with "email → to-dos"

    Claude Pro plus the Gmail connector, 15 minutes to set up, output by tomorrow morning. It's the lowest-barrier, fastest-payoff item of all 22 — make "one flow runs by itself" happen first.

  • Weeks 2–4 · Put it on a schedule

    Upgrade manual triggering to a scheduled routine

    Same flow, but from "I remember to run it" to "it runs itself." Cross that step and you've gone from AI user to workflow builder — all 6 routines above grew exactly this way.

  • Week 5 on · One per month

    Add one per month; delete what doesn't run

    A system is grown, not stacked. Give every new workflow two full weeks — if you're still firefighting manually, you decomposed it wrong. Delete and rebuild. Behind my 22 are at least that many deletions.

There has only ever been one test: "In the past week, which workflow completed itself while I only checked the result?" — if you can answer that, the system is yours.

Real Agent Use Cases

Tools go stale. Workflows don't.

Every one of the 22 workflows on this page gets its own newsletter breakdown: trigger, tools, steps, and the mistakes — all public.
"Execution is permanently undervalued." If you don't plug in one flow tonight, you won't tomorrow night either. That's not motivation — it's observation.

← EP 02 · Multi-agent & Skills　·　EP 03 (you are here)　·　EP 04 · Obsidian Second Brain →
Companion: the 36-creator AI follow list (Chinese) →