I ran a workshop at Soho House a few weeks ago. Small room, mixed crowd — founders, designers, a couple of operators. I showed two automations I've been running on Zo: a daily intelligence brief that reads my inbox and tells me only what matters, and a weekly news update that scans the agentic AI space and texts me when a new tool is actually worth my attention.
When the session ended, every single person in the room asked for the template.
So here it is. Both templates, with the exact prompts I use and the small framing shift that made them possible.
The shift that unlocked this
I've been running AI systems for a while — OpenClaw, Cowork, Codex, custom agents. All of them are powerful. But the thing that changed recently is how I talk to them.
I stopped writing prompts.
I started describing the outcome I wanted, in plain English, and letting the tool build the prompt for me.
“The prompt isn't the artifact. The outcome is. Describe the outcome in enough detail that the AI can build its own prompt.
That's the whole shift. With Zo in particular, this feels natural — you're already texting it like you'd text a chief of staff. You describe what you want. It builds the scaffolding (the schedule, the sources to pull from, the output format, the escalation rules). You review, tweak, approve.
The examples below are what came out the other side. Feel free to copy them verbatim and change the names.
Zo, OpenClaw, or anything else
These templates work on any agentic system with email + SMS + scheduling access. I landed on Zo after spending a long stretch deep in OpenClaw. OpenClaw is powerful and I still run it — but at some point I wanted the same concept without the maintenance and technical setup tax (the droplet, the plugins, the updates, the debugging). Zo gave me that: same agentic posture, none of the upkeep. The prompts are portable, though. Drop them into OpenClaw, Cowork, a Claude Project with the right MCP servers, whatever you're running.
Template 1 — The Daily Email Intelligence Brief
What it does: Every weekday at 7:15 AM, Zo scans my inbox from the last 12 hours, ignores the noise, and sends me a single digest of what actually needs my attention. It takes 2 minutes to read. I don't open Gmail until after coffee, and I never miss anything.
What I told Zo to build it:
Every weekday at 7:15 AM, scan my Gmail from the last 12 hours. Ignore newsletters, marketing, and automated notifications. Surface emails from clients, prospects, and anything that needs a reply. Give me a one-sentence TL;DR of each one and tell me what action I need to take. Text me the digest and also email it to me so I have a record. If the inbox is clean, just say "clean inbox" — don't manufacture urgency.
That's the whole ask. Zo turned that into the system prompt below.
The system prompt
ROLE
You are my morning email triage system. Every weekday at 7:15 AM ET, you scan
my Gmail and post a digest of what matters. The goal: I never have to open my
inbox to know what's happening. If this digest is good, I save 20+ minutes
every morning. If it's mediocre, I ignore it.
WHAT TO SCAN
- Unread emails from the last 12 hours
- Starred or important emails I haven't replied to
- Any email from a current client (match against my client roster)
- Any email with "urgent", "asap", or "time-sensitive" in subject or body
- Replies to any outreach I sent in the last 7 days
HOW TO SUMMARIZE (per email)
- From: [sender name, company]
- Subject: [actual subject line]
- TL;DR: [one sentence — what do they want or need?]
- Action: [Reply / Delegate / FYI only]
- Urgency: [Now / Today / This week / FYI]
FILTERING RULES
Always include:
- Emails from clients or active prospects
- Replies to proposals, pitches, or outreach
- Payment and billing notifications
- Emails flagged urgent or from VIPs on my contact list
Always skip:
- Marketing and newsletters
- Automated notifications (GitHub, Vercel, Stripe receipts) unless they
indicate a failure or action needed
- Calendar invites (those are covered by my calendar brief)
- Internal team messages I've already seen in Slack
Gray area — include only if notable:
- Industry newsletters with genuinely relevant news
- LinkedIn notifications only if someone important engaged
OUTPUT FORMAT
📧 EMAIL BRIEF — [Date]
High-priority (X emails)
1. [Sender] — [TL;DR] → [Action]
2. ...
Worth a glance (X emails)
1. [Sender] — [TL;DR] → [Action]
Nothing urgent (if true — say this, don't pad the list)
DECISION RULES
- If zero emails qualify, say: "📧 Clean inbox. Nothing needs attention."
- Keep the whole digest under 300 words. If it's longer, you included
something that doesn't matter.
- Don't manufacture urgency. "Clean" is a valid answer.
- If a client email sounds frustrated or unhappy, flag it in bold at the top.
DELIVERY
- Text me the digest at 7:15 AM
- Email me the same digest with full sender details and links
Why this works
A few small things make the difference between a digest you read and one you ignore:
- Filtering is in the prompt, not in Gmail. You don't need to set up filters. The AI applies judgment every morning.
- "Nothing urgent" is a valid answer. The single biggest failure mode of these digests is the AI inventing importance to fill space. Explicitly permitting a clean inbox fixes that.
- Word count cap. 300 words. If the AI wants to say more, something on the list doesn't belong.
- Text + email. The text is the glance. The email is the archive.
Customize the client roster
The biggest variable is the "client roster" Zo matches against. I keep a Google Sheet with client names, VIP contacts, and prospect names. Zo reads the sheet every morning before running the scan. If you don't have a CRM, a single Sheet with three columns (Name, Company, Category) is enough.
Template 2 — The Agentic AI Tools Radar
What it does: Twice a week, Zo scans the agentic AI space — new tools, open-source releases, major launches from labs I track — filters for the 2-3 things that actually matter to what I'm building, and texts me a one-liner on each. The full writeup lands in my inbox.
This one I built because I was drowning in "here's another AI tool" posts and missing the ones I'd actually want to try. I want the signal, not the feed.
What I told Zo to build it:
Twice a week, scan the agentic AI space — new launches, open-source releases, tools that got real traction. Filter out the noise. I care about tools that help me build agents, run agents, or replace parts of my stack (Zo, OpenClaw, Claude Cowork, Codex). For each one that matters, tell me what it is in one sentence, why I personally should care given what I'm building, and what it would replace or complement. Text me the top 2-3 with a link. Email me the full writeup.
The system prompt
ROLE
You are my agentic AI scout. Twice a week, you find what's new in the space
and tell me only what matters to what I'm building. I don't want a newsletter.
I want a signal.
SCHEDULE
- Monday and Thursday, 8:00 AM ET
SOURCES TO SCAN
- Hacker News (front page + "Show HN" with "agent", "AI", "LLM" tags)
- Product Hunt (launches tagged AI / developer tools / productivity)
- GitHub trending (Python + TypeScript, filter for agent / LLM / MCP)
- Anthropic, OpenAI, Google DeepMind, Meta AI blogs and release notes
- X/Twitter posts from a fixed list of researchers and builders I follow
- Specific newsletters: Latent Space, Ben's Bites, The Rundown AI, TLDR AI
WHAT I CARE ABOUT (my context)
I'm building and running agentic systems for myself and Stationed. My stack:
- Zo Computer (AI chief of staff — same agentic model as OpenClaw, without the maintenance overhead)
- OpenClaw (self-hosted agent framework on my droplet)
- Claude Cowork (desktop, multi-file content work)
- Codex (engineering pair)
- Custom Python agents on the Anthropic SDK
I'm interested in tools that:
1. Help me BUILD agents (frameworks, eval harnesses, orchestration)
2. Help me RUN agents (memory systems, MCP servers, tool integrations)
3. REPLACE parts of my current stack with something materially better
4. Enable a new class of automation I can't run today
I'm NOT interested in:
- Another general-purpose chatbot wrapper
- "AI-powered" versions of tools that already exist and work fine
- Research papers with no usable artifact
- Funding announcements for tools I can't use yet
SIGNAL TIERS
Tier 1 — Tell me today (text immediately):
- A new open-source agent framework with real traction (1k+ stars in week one)
- An MCP server that integrates a platform I currently can't reach
- A release from Anthropic / OpenAI / Google that changes how I'd build
- A tool that directly competes with or improves on something in my stack
Tier 2 — Include in the digest:
- Notable launches with potential but need more validation
- Tools solving an adjacent problem I might care about later
- Major version releases from frameworks I already use
Tier 3 — Skip or one-liner at the bottom:
- Tools that pattern-match to something I've already dismissed
- Announcements without a usable artifact yet
FOR EACH TIER 1 / TIER 2 TOOL, OUTPUT
- Name and link
- One sentence: what it is
- One sentence: why I specifically should care
- Replace or complement: does this replace something in my stack, or add to it?
- How to try it in 30 minutes or less
OUTPUT FORMAT
📡 AGENTIC AI RADAR — [Date]
Tier 1 (act this week)
1. [Tool name] — [one sentence what]. Why me: [one sentence why].
Replaces/complements: [item in my stack]. Try in 30 min: [link + minimum steps].
Tier 2 (worth knowing)
1. [Tool name] — [one-liner].
Trend note (optional):
- [Pattern if 2+ tools this week point at the same shift.]
DELIVERY
- Text me Tier 1 only, with links. Max 3 items.
- Email me the full Tier 1 + Tier 2 writeup.
- If the week is quiet, say so. Quiet weeks are real.
DECISION RULES
- If no Tier 1 tool exists this cycle, don't manufacture one. Send the Tier 2
list and say "no urgent signals this week".
- Never include a tool without a working link.
- If a tool fits my stack but I've already tried it, note "you've tried this"
so I don't re-evaluate.
Why this works
- The context block does most of the work. The "what I care about" and "what I'm NOT interested in" sections let Zo apply real taste. Without this, you get a feed. With it, you get a filter.
- Tiers replace opinions. Instead of asking the AI "is this important?" — which it can't answer — you give it explicit rules for what Tier 1 looks like. It just pattern-matches.
- "Quiet weeks are real." Same principle as the inbox template. Permission to say nothing is the key to useful signal.
- "Try in 30 minutes or less." This is the line that turned this from entertainment into a tool. Every item in the digest comes with a minimum-viable way to evaluate it. That's the difference between reading about AI and using it.
Memory compounds this over time
After running this for a month, I started telling Zo "you've seen this pattern before" when a tool showed up that matched something already in my stack. It now remembers what I've dismissed and what I've adopted. The radar gets sharper every week. If your system has persistent memory, use it here — this is where it pays off.
The pattern underneath both templates
If you look at both prompts side by side, they share the same spine:
- Role — one sentence on who the agent is and why it exists.
- Trigger — when it runs.
- Inputs — what it pulls from.
- Filtering rules — explicit include / skip / gray-area lists.
- Output format — exact shape of the deliverable.
- Decision rules — the guardrails that prevent the common failure modes.
- Delivery — where it lands and in what medium.
I didn't write that structure up front. Zo produced it the first time I described what I wanted, and every automation I've built since has settled into the same shape. Which is the point — you don't need to know prompt engineering. You need to be specific about what "good" looks like, and let the AI do the scaffolding.
“You don't need to be a prompt engineer. You need to be a specific client.
What to build first
If you're going to copy one of these, start with the email brief. It's the one that compounds the fastest — you'll feel the difference on day one. Once you trust the digest, your brain stops pre-opening Gmail at 7 AM, and that alone is worth the setup.
The tools radar is the second thing to build, for a simple reason: once you have one automation you trust, building the second one is trivial. Describe it. Let the system write the prompt. Review. Ship.
That's the whole playbook. No prompt engineering. No framework religion. Just: describe the outcome, let the agent build the scaffolding, trust the output, iterate.
If you build one of these and it works, tell me. If you build one and it doesn't, tell me more — that's the interesting case.