I don't have a productivity system anymore. I have a team.
Five AI employees running on my own Mac. They're online 24/7, they talk to me on Telegram, they coordinate through a shared backlog and sprint board, and they each own jobs I used to do by hand. I review their work for about 30 minutes a day. The rest happens while I'm asleep, in meetings, or shooting content.
This is Issue #04 of the dispatch, the workforce edition. A breakdown of the five concrete jobs my AI team did this week, with the schedules each runs on, what they hand me, and the rough time each one saved. No theory. Just what's actually running.
Job 01: Inbox triage (every weekday, 9:00 AM)
The employee: Quill, my Google Workspace manager.
What she does: Reads every unread email in Gmail, drafts replies in my voice, files the noise.
By 9:30 AM, my inbox has one sheet open in front of me: a list of drafts to review. The drafts cover the actual work: contract redlines, partnership terms, warm intros, calendar requests. The noise (invoices, generic outreach, "circling back" follow-ups) was filed without ever surfacing.
A real example from this week:
Re: Q3 partnership terms "Thanks for the deck — drafted a yes with two ask-back questions. Review →"
Re: contract redlines "Drafted response on 4 of 5 redlines. One needs your call."
Re: warm intro request "Drafted warm intro + calendar link. Awaiting your sign-off."
I tap accept on the ones that read right, edit the ones that don't, and send. The 38-unread morning becomes a 6-decision morning. Time saved this week: ~4 hours.
How it's wired: Quill has the Gmail MCP server and a Claude Code routine that fires every weekday at 9. The routine prompt is a 200-word brief I wrote once: how I write, who matters, what to file, what to flag. She reads it, plus my boss profile (who I am, what I'm working on, who's important right now), and after eight weeks of feedback produces drafts that sound like me.
Job 02: Trending scan (every day, 7:00 AM)
The employee: Nova, my content strategist.
What she does: Pulls what's blowing up on GitHub, X, and Hacker News overnight. Drafts a video brief I can shoot the same morning.
The output that landed in my inbox this week:
Today's report — scanned 14 sources · 06:55 AM
- claude-context (zilliztech) — Semantic code search MCP · 10.6k stars
- pi-mono (badlogic) — Agent toolkit · unified LLM API
- ml-intern (huggingface) — Open ML engineer that reads papers
- TradingAgents (TauricResearch) — Multi-agent finance firm · 62.6k stars
- Pixelle-Video (AIDC-AI) — End-to-end video gen pipeline
Pattern this week: not "another agent framework." Coding-context MCP, ML agent that reads papers, multi-agent firm in code. The era is multi-agent specialists.
That feed became the GitHub roundup blog post you're probably here from. Nova does the boring half of that post: finding the repos, sorting by traction, pulling the descriptions, spotting the pattern. I do the interesting half: picking which to feature and writing the actual takes.
Without Nova I was spending 90 minutes every morning scrolling Trendshift, OSSInsight, and HN. Time saved this week: ~5 hours. And the videos hit the trends a day earlier than my competitors who do it by hand.
Job 03: Branded proposals (on demand, ~12 minutes)
The employee: Atlas, my proposal writer.
What he does: Takes call notes and a project scope, returns a fully branded PDF proposal with cover, scope, timeline, pricing, and T&Cs.
I dropped this brief in Telegram at 3:14pm yesterday:
"Discovery call done. Brand system + 8-page web build + CMS migration. 8-week timeline. 50/25/25 milestones. 2 revision rounds, 30-day post-launch support."
At 3:26pm I had a 14-page PDF in my inbox: cover with the prospect's name and logo, executive summary in my voice, scope written from my call notes, a Gantt-style timeline, three pricing tiers with milestone breakdowns, T&Cs from my legal template. Branded, formatted, ready to send.
What used to take me an evening of moving Pages around takes 12 minutes of one prompt. Time saved per proposal: ~3 hours. I sent four this week.
How it's wired: Atlas owns a skill called make-proposal that takes a brief, fills my Typst template, generates the PDF, and pushes it to my CDN. The template is brand-locked (typography, colors, layout) so every proposal looks identical no matter how messy my brief was.
Job 04: VPS / DevOps (24/7 ops)
The employee: Rack, my full-stack DevOps engineer.
What he does: Owns the servers. Uploads files, edits configs, deploys containers, syncs to the CDN, schedules social posts, monitors logs.
This is the employee I see the least and trust the most. He shipped 23 ops this week and zero pages reached me. A snapshot of his weekly log:
rack@glitch · vps · this week
✓ vps · pulled latest · restarted
✓ media/asset.png → cdn
✓ proxy/config · added staging route
✓ app v2.4 → registry
✓ --schedule "tuesday tip" → social
✓ app/logs · 0 errors
✓ uploads → cdn
all containers · all sites · all green
✓ 23 ops this week · 0 pages to me
The reason this works without me losing sleep: every action Rack takes goes through a Claude Code hook that requires explicit confirmation for anything destructive (deletes, force-pushes, prod deploys). The 23 ops he ships are the safe ones. If something needs my call, he Telegrams me before doing it.
Time saved this week: ~6 hours of small DevOps tasks. More importantly: zero context-switches. I never had to leave the thing I was doing to "just quickly" deploy a fix.
Job 05: Vault sync (after every meeting, ~1 minute)
The employee: A bundled skill that runs across all of them.
What it does: Every call I take, every project spec I'm sent, every clip I save gets pulled into my Obsidian vault using Karpathy's LLM Wiki method. Cross-linked, summarized, integrated.
This week's writes to the vault, surfaced in Sunday's /digest:
Today's writes to Wiki/
acme-account.md— 12 lines added · contract notes from Wed callpricing-strategy.md— restructured · 3 new comps from Stripe intel2026-q1-recap.md— stitched 4 weekly digestsstripe-vs-paddle.md— new file · 2 sources cited
The result: every conversation I have feeds the brain. When a prospect asks me about pricing, I'm not pulling from memory. I'm pulling from a wiki page that's been getting updated for six months by the agent. The synthesis happens once. The vault gets smarter every day.
Time saved this week: ~3 hours. The bigger win is qualitative: I stopped forgetting things from calls. That's worth more than the 3 hours.
The setup, end-to-end
In case it's useful: the whole thing runs on the Glitch Workforce, an open-source AI workforce framework that lives on my own Mac. Every employee is a Claude Code instance with their own:
- Identity: a
CLAUDE.mdfile that defines who they are, what they own, and what they're not allowed to touch - Skills: markdown playbooks for the recurring jobs they do
- Schedules: Claude Code routines for "every weekday at 9 AM" type work
- Telegram bot: so I can talk to each of them from my phone, not just my laptop
- Shared backlog: the team coordinates through a single sprint board so I can see all 5 employees' work in one Command Deck
You don't write any of this from scratch. You install the Workforce, type "I want to hire someone who does X" to Professor Glitch on Telegram, and the system creates the employee (desk, identity, bot, skills, schedules) in about 5 minutes.
Total time saved across the five jobs this week: roughly 21 hours. Roughly half a workweek. While I shoot content, take meetings, or sleep.
The payoff (what actually changed)
The honest answer to "what did I get from this?" isn't the 21 hours. The 21 hours is nice. The bigger thing is the shape of my week.
I used to context-switch 30+ times a day. Inbox → Slack → invoice → "let me just deploy this fix" → "let me draft this proposal" → back to writing → check trending → meeting → forgot what I was doing. That's not a productivity system. That's a tax on every block of focus you try to keep.
Now I have one job per day: decide. I show up, I review the drafts, I approve the ops, I edit the proposal copy, I tell the trending scanner which repo I want a video on. Every employee owns their lane. I own the call.
"I don't have a productivity system anymore. I have a team."
That's the part nobody tells you about a real AI workforce. It's not "I do my work faster." It's "I do different work." The tedious work isn't sped up. It's gone. What's left is the work only you can do.
How to start (without rebuilding your life)
If this resonates and you don't want to wait three months to feel a result:
- Pick the one job that drains you most. Inbox? Content trending? Bookkeeping? Proposals? Don't try all five. Pick one.
- Install Glitch Workforce. It's open-source, runs on your Mac, takes 10 minutes to set up.
- Hire one employee for that one job. Type to Professor Glitch on Telegram: "I want someone who handles X every weekday at Y." He'll walk you through the rest.
- Use it for 2 weeks before adding a second one. Tune the prompts, the schedule, the voice. By week 3 you'll have a real second hire idea.
The mistake I see most people make: trying to build the whole team on day one. Don't. Build one employee, prove it works, then expand. By month 3 you'll have 4–5. By month 6 you'll have a system you can't imagine working without.
Want the structured path?
Two things, depending on where you are:
- Join the community for structured paths, hands-on labs, and the Thinker's Mind episodes on how to think about AI before you reach for a tool. 130+ builders shipping real AI workforces inside.
- Follow @pro.glitch on TikTok for the free daily breakdowns on how each piece works.
The hardest part of getting an AI workforce running isn't the technical setup. It's the leap of faith required to actually delegate something. Start with the smallest possible job. Watch it work for two weeks. Then start letting go of the next one.
The first hire is the hardest. The fifth one feels obvious.