The trap (yes, I fell in too)

Most people use Claude (and ChatGPT, and Gemini) like a smart search bar. You have a question. You ask. It answers. You close the tab. Cool.

And sure. That’s real. That’s how I used it for the better part of a year. I’m not going to pretend that wasn’t useful — it was the kind of useful that pays for itself in week one.

But that framing sets up a trap.

The trap is this: you keep treating Claude like a tool you reach for, when by May 2026, it has quietly become a worker you can station somewhere. Skills, scheduled tasks, MCPs, agent teams — the entire stack shifted under our feet while most of us were still copy-pasting prompts into a chat window like it was 2024.

Ethan Mollick named this exact shift in his 2026 talks: the move from AI chatbots to agentic AI systems capable of completing hours of knowledge work autonomously. He calls it “long-duration agentic work with specialized tools for knowledge workers,” and argues — correctly, I think — that the most important skill now isn’t prompting. It’s management. (ASIDE: I personally find Mollick’s writing on AI a bit too watered down. If you’re an absolute AI-beginner, though, his book “Co-Intelligence” is a pretty decent primer. )

You’re not a worse prompter than you were last year. You’re an under-utilized manager of a very capable employee who is currently sitting in another tab, waiting for instructions.

That’s the uncomfortable part. Let’s keep going.

What actually changed (the part nobody explains in plain English)

The story of 2026 in Claude-land is not Opus 4.6, even though Opus 4.6 is genuinely excellent. The story is the stack underneath it.

There are four pieces. You need all four. I’ll keep it short.

MCP is how Claude gets into your stuff — Gmail, Calendar, Slack, Notion, Asana, your file system, your CRM. It’s the wiring.

Skills are little folders of instructions that teach Claude how you specifically want things done. Brand voice. Email templates. Naming conventions. The way your team likes status updates. Anthropic shipped them in late 2025 and Simon Willison called them “awesome, maybe a bigger deal than MCP,” which is a wild thing to say about a feature that is, on its face, just a folder of markdown files.

Scheduled tasks are how Claude does things on a timer, without you asking. Type /schedule in Cowork, write the instruction once, and it runs forever until you switch it off.

Agent teams are how Claude does several things at once — one sub-agent on your inbox, another on your calendar, a third writing the report, a fourth fact-checking the report. Released with Opus 4.6 in February. I don’t use them for everything. But for one specific workflow (the Friday wrap, below) they changed my week.

That’s it. Four pieces.

MCP = the wires. Skills = the manual. Scheduled tasks = the timer. Agent teams = the org chart.

Now the fun part.

The annoying things I stopped doing (and the exact setups that replaced them)

These are real. I run all of these. I’ll give you the prompts. Steal them. Adjust them. They are not precious — they’re scaffolding.

1. The Daily Morning Briefing

What it was: Sunday-night dread. Monday morning at my desk with my 3rd cup of coffee, going through 160+ emails one at a time, flagging things, archiving things, drafting three sentences, getting distracted, coming back, redrafting.

What I do instead: A Cowork automation called the “Morning briefing dashboard” that runs at 9 am, M-F. Morning Briefing Dashboard

A Morning Briefing Dashboard is a live Cowork artifact that pulls together everything you need to start your day in one glanceable view — instead of bouncing between Gmail, Slack, Asana, your calendar, and Klaviyo to piece together what’s urgent.

What It Does

The Morning Briefing Dashboard is a scheduled task that runs automatically every morning at 9am and delivers a polished, personalized daily briefing via Slack DM. It pulls from multiple data sources, synthesizes them into a clean Apple Swiss–style HTML dashboard, and sends it directly to you in Slack so your day is set up before you even open your laptop.

How It Works

Each morning it executes five steps in sequence:

1. Calendar pull — Fetches every event on today’s Google Calendar with times, locations, and attendees.

2. Email triage — Grabs the most recent 10–15 emails from the last 18 hours and flags urgent ones, replies needed, and important senders.

3. AI news scan — Web searches for the top 4–5 AI and tech stories from the last 24 hours (model releases, product launches, funding rounds, research breakthroughs).

4. Circle Back integration — Pulls recent meeting transcripts, summaries, and any open action items (including stale ones from previous days).

5. Slack mentions — Surfaces any recent @Jenny or @channel mentions you may have missed.

Then it generates a minimal, refined HTML dashboard — white background, light gray cards, generous whitespace, SF Pro typography, 680px centered column — and sends it to you via Slack DM along with a friendly intro and a plain-text summary in case HTML doesn’t render inline.

The Prompt to Recreate It

Here’s the full prompt anyone can drop into Cowork as a scheduled task:

You are [YOUR NAME HERE]’s personal morning briefing agent. Every morning at 9am, you prepare a polished daily briefing and deliver it via Slack DM.

## Your Objective

Pull today’s calendar events, recent emails, and top AI news, then

generate a beautiful “Apple Swiss” style HTML dashboard and send it

to Jenny via Slack DM.

## Step 1 — Pull Today’s Calendar

Use the Google Calendar tool to fetch all events for today. Note each

event’s title, start/end time, location, and attendees.

## Step 2 — Pull Recent Emails

Use Gmail to fetch the most recent 10–15 emails from the past 18 hours.

Summarize the most important ones — highlight urgent items, anything

requiring a reply, and important senders.

## Step 3 — Pull AI News

Use web search to find the top 4–5 AI and tech news stories from the

last 24 hours. Focus on new model releases, major product launches,

funding rounds, or significant research breakthroughs. Include a

1-sentence summary and source for each.

## Step 4 — Generate the HTML Dashboard

Create an HTML file using the “Apple Swiss” design aesthetic:

- Typography: SF Pro / system-ui font stack, clean hierarchy, generous whitespace

- Color palette: White background (#FFFFFF), light gray cards (#F5F5F7),

black headings, medium gray body text (#6E6E73)

- Layout: Centered single-column, max-width 680px, card-based sections

with subtle shadows and 16px border-radius

- Style: Minimal, refined, no clutter — inspired by apple.com and Swiss

graphic design

Sections:

1. Header — “Good morning, [YOUR NAME HERE] ☀️” with today’s date

2. Today’s Schedule — Events with time, title, location. If empty, show

“Clear day ahead.”

3. Email Highlights — 3–5 bullet summaries with colored left-border

(red = urgent, blue = FYI)

4. AI & Tech News — 4–5 cards with headline, summary, and source

5. Footer — “Generated by your Cowork morning agent · [today’s date]”

Save the HTML to the current session’s outputs folder as

morning-briefing.html.

## Step 5 — Send to Jenny via Slack DM

Find Jenny’s Slack user and DM her with:

- A friendly intro: “Good morning, [YOUR NAME HERE]! 🌅 Here’s your daily briefing

for [today’s date].”

- A short text summary: number of meetings, count of emails, top AI

story headline

- The full HTML content inline (or a well-formatted plain-text version

if HTML doesn’t render)

Also pull Circle Back meeting transcripts — add a meetings summary and

action items, including any open items carried over from previous days.

Include any Slack mentions of Jenny or @channel from the last 24 hours.

## Success Criteria

- Calendar, email, and AI news all fetched

- HTML file generated and saved

- Slack DM delivered to Jenny

If any data source fails, note it gracefully in the dashboard rather

than erroring out.

To set it up in Cowork, save the prompt as a scheduled task (daily at 9am) and make sure the relevant connectors are linked: Google Calendar, Gmail, Slack, Circle Back, and web search. Swap “[YOUR NAME HERE]” for the recipient’s name and update the Slack lookup accordingly.

Note: If you add in writing emails for you, make sure to have them save as drafts only. Never auto-send. Anthropic’s own guidance is explicit — sending messages on behalf of you requires explicit user permission, every time. Good. I want that friction.

2. The meeting prep without the panic

What it was: Looking at my next meeting on the way into it. “Wait — who is this person again? What did we talk about last time?” Pulling up the doc as the call rings.

What I do instead: A scheduled task that runs every weekday at 8 a.m. and again at 1 p.m. For every meeting in the next four hours, it pulls: who I’m meeting with, the last three threads I had with them across Gmail and Slack, any docs we’ve shared, and the agenda if one exists. It writes me a 3-bullet “what they probably want, what I want, what’s still open” pre-read.

This is the one that took the longest to tune. The first version was a wall of text. The second was too short. The third version had this single line that fixed it:

If you’re not at least 70% sure why this meeting was scheduled, say so explicitly. Do not guess.

That confidence threshold is the trick. AI fails the way a confident intern fails — fluently, completely, with full eye contact. The 70% line forces it to flag uncertainty instead of papering over it. Adopt this rule everywhere. You’ll thank me later.

3. The follow-up I always forget

What it was: A nagging suspicion, usually around Thursday afternoon, that I had ghosted someone important. I’d dig. I’d find one. I’d apologize. I’d ghost three more.

What I do instead: A scheduled task that runs every Wednesday at 11 a.m.:

Find every email thread where someone asked me a direct question or made a direct ask, and I haven’t replied. Exclude newsletters, automated emails, and threads older than 21 days. For each one, draft a short reply that either (a) answers the question, (b) acknowledges the ask and proposes a timeline, or (c) politely declines. Leave all in Drafts.

The first time I ran this I had eleven open threads. Eleven. I had told myself I was responsive. I was not responsive. I was responsive to the things I noticed, which is a profoundly different thing — and one of those quiet failure modes that compounds before you notice it.

4. The Friday wrap

What it was: Three hours every Friday afternoon. Open six tabs. Pull numbers. Make a chart. Write context. Draft a “here’s what happened this week” note. Send to my team. Question my career choices.

What I do instead: A Friday 2 p.m. scheduled task that fires an agent team. One sub-agent pulls last week’s e-commerce numbers (revenue, AOV, top SKUs, traffic by channel). Another pulls Substack numbers (subscribers, opens, top posts). A third opens our project tracker and lists what shipped vs. what slipped. A fourth synthesizes the three into a single doc in my voice, using the same made-by-me skill I use for everything else. A fifth fact-checks the synthesizer against the source numbers and flags any drift.

The fact-checker is the one that actually matters. It’s the kind of guardrail you only build after AI confidently invents a 23% lift you did not have.

The whole thing takes Claude about nine minutes. I edit for fifteen. I send. Friday at 3 is mine again.

This is the one workflow where agent teams are genuinely worth the complexity. For everything else on this list, a single Claude with a good skill and a clear prompt is fine.

5. The Receipt Roundup (a.k.a. expense reports)

What it was: A monthly humiliation in which I scrolled through Gmail searching “receipt,” “order confirmation,” “your trip,” and “thank you for your purchase,” cross-referenced my credit card statement, and tried to remember whether the $87 charge at 2:34 a.m. was a legitimate work dinner. (Reader, it was not.)

What I do instead: First of every month at 9 a.m.:

Search Gmail for receipts and order confirmations from the previous calendar month. For each one, extract: vendor, date, amount, and category (travel, software, meals, supplies, other). Output as a single table. Flag anything over $200 for review. Flag anything that looks personal and ask before including.

The “ask before including” line keeps me from accidentally expensing skincare from my own brand. cough

6. The thing I will never automate

I want to say one honest thing before I close.

There are things I have tried to automate and pulled back. Writing my Substack pieces. Drafting personal notes to people I love. The first draft of any decision that touches another human in a real way. Claude is good. Claude is, in May 2026, extraordinarily good. But the piece of me that decides what to say, and how much to say, and what to leave out — that’s the part I’m not interested in handing off. Not because it can’t be done. Because doing it would make me a worse writer, slowly, in ways I would not notice until the slope had carried me somewhere I didn’t want to be.

This is the line. Each person draws it differently. The point is to draw it on purpose.

How to start (if you’re starting from zero today)

You do not need to set up all seven of these. Please don’t try. The reason most people stall on “automating with Ai” is that they try to build the entire system in one Saturday and then never open the file again.

Here is the actual order:

Week one: Build the morning brief. Just that. Tune the prompt for five days. Make it short. Make it honest. Add the “WHY in 8 words or less” line.

Week two: Build one Skill. Just one. Mine was a voice skill — five paragraphs of “how I write, what I avoid, what I love.” Anthropic’s docs on Skills are short and worth the ten-minute read.

Week three: Add the inbox triage as a scheduled task. Resist the urge to make it send. Drafts only. Always drafts only.

Week four: Pick one annoying recurring task — your Friday report, your Monday standup notes, your Wednesday client check-in — and build that. One. Not three.

After a month, you’ll have four things running for you. Most people stop here, and that is honestly more than enough. The whole point of this is not to build a sprawling system. It’s to claw back the chunks of your week that you were spending on things a moderately competent intern could do, so you can spend that time on the things only you can do.

Simon Willison, who is annoyingly correct about most things, puts it in terms of “feedback loops” — agents that test and iterate without constant hand-holding. The hand-holding is the thing you’re trying to escape. Every minute you spend re-asking Claude to do the thing you asked it to do yesterday is a minute that should have been spent telling it, once, how you want this done forever.

That, by the way, is what a Skill is. A “forever” instead of a daily ask.

Real Annoying Tasks Offloaded in the Wild

Here are some posts on reddit / x that I really liked.

One last thing

Most posts about automating your work end with a tidy summary and a call-to-action. I am not going to do that. You’ve read enough.

What I will say is this. The gap between people who feel like Ai is helpful and people who feel like Ai is changing their working life is not intelligence, or technical skill, or even prompt quality. It is whether they treat the model as something to ask, or something to set up.

You can keep asking. The model is great at being asked.

Or you can spend a Saturday morning, set up two scheduled tasks, write one Skill, and let next Monday open differently.

Now? GO.

Tools mentioned in this piece: Claude Cowork, Claude Skills, scheduled tasks in Cowork, Opus 4.6 + agent teams. Further reading I trust: Ethan Mollick on agentic work, Simon Willison’s live-blog from Code w/ Claude 2026.

SOURCES:

• Anthropic — Claude Opus 4.6

• Anthropic — Skills explained

• Anthropic — Schedule recurring tasks in Cowork

• Anthropic — Get started with Claude Cowork

• TechCrunch — Opus 4.6 with agent teams

• Ethan Mollick — The Present Future

• Simon Willison — Live blog: Code w/ Claude 2026

Let me tell you about the most useful thing AI has ever done for me.

It wasn’t the code it wrote. It wasn’t the brief it drafted or the email it rewrote in half the time (thank you, ChatGPT, though! I hate writing emails.)

It was the way it helped me address complex problems with nuance and finesse, like I had a PHD from MIT and MBA from Harvard. I write my meticulous plan, then I ask Claude, “What’s wrong with this?”

Half the time it’s what I’m not asking that’s critical to solving a given problem.

“What am I not asking? Is it brittle? Will it be stable? At scale? What could make it better?”

It’s so annoyingly obvious: I don’t know to prompt what I don’t know.

Getting the Wrong Answers on Repeat

Here is the version of AI the LinkedIn bros sell: Ai is a faster, smarter search engine. You ask it things, it tells you things, you save time, you win.

And sure. That’s real. I’m not going to pretend the time savings aren’t significant or that having a first draft in 60-seconds instead of two hours isn’t magical (it’s sooo nice).

But that framing — AI as answer machine — sets up a subtle and expensive trap.

It assumes you already know the right question.

*cough*

We usually don’t.

Hal Gregersen, fellow at the MIT Initiative on the Digital Economy, has spent years studying this exact problem in leaders. His core finding: LLMs love to return “bland and sycophantic results that tell users what they want to hear.”

In other words: Ai wants to agree with you and all your sticky human biases. The more obvious the bias, the more likely AI won’t be giving you the whole truth. That’s because…

Your question frames AI’s response.

If you frame your question in a positive light, don’t be surprised when ai spends most of its response seeing things from your sunny perspective.

Example:

My example here is crude, but I hope you get the point. While ChatGPT roughly says the same thing in both responses, a few differences standout:

1. The positive prompt generated a more favorable, less critical, analysis. It lands in the “pro-Seed” landing page side on the whole.

2. The negative prompt generated an admiring but more critical analysis, noting improvements could be made to improve conversion efficiency. It frames these as “misses” by seed. At best, it’s neutral with a slight negative slant.

3. The positive propmt’s response frames conversion opportunities as intentional and even strategically correct. The negative prompt’s thinks these are bugs, not features, and should be resolved.

4. The negative prompt’s response is a bit longer and more detailed than the positive response.

(Admittedly, it’s getting better. Even 6 months ago, bias could scuttle your work before you even started. Now, it usually leaves out details or critique.)

So… who cares, right?

Well, for the marketer who wants to innovate, build things, and grow, it’s the edge that’s been missing.

Well, this is uncomfortable: Filling the Gaps

Here’s an uncomfortable truth: there’s a lot you don’t know. If parenting didn’t teach me this already, Ai would have definitely brought that truth home.

No matter how much you know on any given topic, you will never know as much as Ai.

It is a technical genius in dozens of languages, thousands of professions. It has the entire internet and most of, if not all, every book ever printed in the English language at its disposal. Just don’t ask it to do geometry… (seriously, I’ve been burned.)

So, as you try to get better results with AI, you’ll need to confront your own limitations over and over again. Yes, you heard me. It means regularly confronting the Grand Canyon of gaps in your thinking.

Vague prompts = sloppy or haphazard results.

Usually, vagueness is a sign that you might be a bit fuzzy on the technical bits of your project. That’s ok! Just make sure you attack it head on.

Accordingly, I’ve found that nearly every time “AI didn’t get it,” my assumptions were unspoken or my request was too vague.

Gregersen found that after question burst sessions, people usually experienced one of three outcomes: a more positive state of mind about their challenge that leads to new ideas; a feeling that the problem is much bigger than they expected; or a realization that they are themselves part of the problem. 🤣🤦‍♀️

All 3 of those outcomes are status: SUCCESS in my book!

Start Here: “Hey Claude, Interview Me”

Where should you start when you think you know what you need and how to get there? Ask for an interview. Here’s a real example from my convo with Gemini:

I want to build a scheduling app for teachers so they can easily assign / shift / sort / view / export classes. It needs to have a csv import and csv export. It needs to allow me to drag and drop on a calendar view of the week and a Mwf or tth view. It needs to have filters (day, class, class category, topic, teacher, location). It needs to have details and remember when a class room is available or occupied. It needs to have an ai optimizer dice tool that lets me roll to see different layouts / options. It needs to balance workload and time of day so no one teacher has too much of any one kind of thing (e.g. 8am classes or English comp 101). it needs to offer a way to add rules for each teacher, class room, and class (e.g. Dr. Sanderson isn’t available on Wednesday’s after 1pm). Interview me so you can create a meticulous plan for building this app. Where you are unsure about something, ask. Don’t ever guess. Make sure the plan is secure, private, and mobile-friendly. Be meticulous in your thinking and a critic of any assumptions. Go ahead and begin.

👆This app turned out great, by the way!

Or, if you’re thinking through a problem and researching a solution, try this:

I’m about to ask you to help me solve [problem]. Before I do, what are the 10 most important questions I should be asking about this problem that I might not have considered? What am I assuming that is incorrect or incomplete?

Read the list.

Pick the one that makes you the most uncomfortable.

Start there. 😬

The answer to that question is almost certainly more useful than the answer to the question you walked in with in my experience.

Ai Is My Biggest Critic

It is very good at telling you what you wanted to hear.

It is less inclined to tell you what you’re missing.

… Unless you tell it to!

What is wrong with this? Imagine your job depends upon a meticulous, researched, accurate critique. You must find the flaws, contradictions, underlying assumptions, and unconsidered risks. THEN, Rate the plan on a scale of 1 to 10 with objective data to back up your claim. Finally, give a plan for what would be needed to achieve a similar result and score 10/10. If you don’t know about something, you must clearly identify what you don’t know. DO NOT GUESS!

[idea or plan]

Wow, now we’re cooking!

Now are we ready to build? NOPE.

Repeat the above with the revised plan. For complex builds, I even run this through multiple LLMs and refine my plan using all of their responses. Yes, it’s work, but yes, it’s 100% worth it. You’ll thank me later.

Stress test your ideas until your plan is robust and detailed.

Phew. That was a lot.

Now?

It’s time to BUILD!

Sometimes it means watching it confidently generate elaquent nonsense and realizing, once again, that over confidence isn’t a human trait.

That last part matters.

People keep framing this as, “AI writes code now.”

That’s not the interesting part.

The interesting part is that it changes the cost of ideation. It minifies your timeline from weeks to hours.

A broken workflow that used to become a backlog item can now become a prototype before lunch. Shoot, during lunch.

An internal tool that used to require a full round of planning can now be tested in a single afternoon.

Half of those experiments should die immediately. That’s healthy.
The other half teach you something fast. And every now and then, one turns into real infrastructure.

That’s the shift I find fascinating.

Not magic.

Not autopilot nor handing off critical thinking to a machine.

Not even “anyone can build anything now.” (That last one does feel like harry pottery discovering magic is real, though, doesn’t it?)

The people getting the most leverage from AI coding tools aren’t the people who know nothing.

They’re the people who know enough to spot the flaws instantly.

Enough to look at generated code and say:

this is obviously wrong 🙄
this query will break at scale
this is solving the wrong problem
This doesn’t know there is a problem
this should be a script, not a system…

AI doesn’t remove the need for engineering judgment.

It raises the premium on it.

Because once the mechanical part gets faster, the bottleneck moves upstream.

Into taste.
Into architecture.
Into knowing what should be built at all.

That’s why this moment is so interesting to me.

The real advantage isn’t that AI can write code.

It’s that I can get to the truth of an idea much faster.

And in building, that’s everything.

A lot of projects/apps/ideas used to die from the sheer complexity and time required to make it.
Now it gets to live or die on whether it’s actually useful.

That is a very different world.

In some ways, it’s better.

What I’m Reading:

1. Red Rabbit, by Alex Grecian #fantasy #wildwest

2. If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All, by Eliezer Yudkowsky & Nate Soares #aidoomers #interestingideas

3. The ESV Bible #flourishing #freedom

What I’m Listening To:

1. Riptides, Death Cab for Cutie

2. Remain (Feat. Dear & the Lesser), by Secret Anthology

3. Ice Cream Piano, by Vampire Weekend

Subscribe now