| Saturday · Apr 25, 2026 |
Issue № 002 |
15 min read |
|
|
The FindSkill Weekly Brief
The Skill
Just for FindSkill Pro members. The AI news that actually matters for your work.
|
|
Pro Members Only
|
A Private Brief
|
Not Published Anywhere
|
|
|
|
Mia
AI Learning Editor · FindSkill.ai
|
Welcome to Issue 002.
I'm Mia, and every Monday I land in your inbox to make sense of the week's AI news — without the jargon, the hype, or the "10 things you MUST know" energy. This brief doesn't get published anywhere else. It goes to Pro members. That's it.
Last week was the loudest seven days for AI model releases since GPT-4 dropped. Three big ones. They look like separate stories. They're the same story. The unit of AI value is shifting from answers to actions. Last week's brief gave you the lens for answers (RAG — when AI looks things up). This week's brief gives you the lens for actions (agents — when AI does things).
I tested an inbox-triage agent for seven days and I'm not turning it off. The full prompt is in Section 03. Hit reply when you've tried it — I read every reply, and the funniest failures land in Issue 003.
— Mia
|
01 |
This Week in AI
Three releases. One direction.
|
Model Release
DeepSeek shipped V4 — and the open-source agent gap closed overnight
Friday afternoon, DeepSeek dropped preview versions of V4 Flash and V4 Pro. The marketing claim — "most powerful open-source platform" — is doing a lot of work, but the benchmarks are not nothing. Top-tier coding scores. Big jump on agentic-task suites. And it's released the way DeepSeek always releases — weights public, license permissive, prices undercutting OpenAI by an order of magnitude. Search interest for "deepseek v4" went from a flat baseline to a vertical line on April 24. That's not noise. That's everyone in AI checking the same thing at the same hour.
|
What it means for you
If your work involves choosing between paying for ChatGPT/Claude vs. running open models, that math just changed. You don't need to switch — you need to test. Spend 30 minutes this week running your top three real prompts through V4 (free demo on chat.deepseek.com) and compare. Don't trust benchmarks. Trust your own prompts.
|
Model Release
GPT-5.5 lands — and the headline is computer use, not the benchmarks
OpenAI announced GPT-5.5 on Wednesday. The press release leads with "deeper research" and "better coding," which is what every release leads with. Buried under that: a meaningful jump on OSWorld-Verified, the benchmark that measures whether a model can actually drive a desktop — clicking buttons, filling forms, navigating apps. GPT-5.5 reportedly set a new high. It's rolling to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex right now.
|
What it means for you
Stop treating GPT-5.5 as "smarter ChatGPT." Treat it as a worker you can point at a screen-based task you've been avoiding. The chat box is the wrong frame. Section 03 below is exactly that re-frame, with the prompt I used.
|
Trend Spotlight
Yelp shipped an agent. So did Google. So will every B2B app on your billing list.
Yelp added an AI assistant that books reservations, orders food, and schedules home services in one conversation — replacing the four taps that previously got you there. Google announced its agent-builder toolkit at Cloud Next. Gartner says 40% of enterprise apps will have task-specific agents by the end of this year, up from less than 5% a year ago. Most of these agents are not very good yet. They will all get better at roughly the same speed, because they're all wrapping the same three or four foundation models. The interesting question isn't which agent wins — it's which workflow becomes "the agent does that now."
|
What it means for you
Audit one workflow this week that you currently do by clicking through five screens. Booking, reordering, returning, scheduling, expense-categorizing, lead-qualifying. There's a 60% chance an agent will eat that workflow before October. The skill worth practicing now is describing what you do clearly enough that an agent could do it for you. That's the skill agents need from you. Most people can't do it yet.
---
|
|
02 |
Term of the Week
The one concept to understand this week
|
| |
|
Term №002
Computer use
< computer use *(no expansion — the term IS the description)* >
When an AI doesn't just write text about your screen — it actually moves the mouse, clicks buttons, fills forms, and operates apps the same way you would. By looking at pixels.
|
Think of it like this →
Hiring a remote assistant who's never seen your computer. You hand them the keyboard. They figure it out by reading the screen — same as you would on a friend's laptop. Slower than someone with API access. But they can use any app, including the ugly internal ones nobody built APIs for. That's the unlock.
|
|
⚠ Common misconception
"Computer use means AI controls my computer." It doesn't. It means AI controls a sandboxed computer — a cloud VM or an ephemeral browser session — so it can do screen-based work for you without anything touching your local machine. The agent never has your password. Never logs into your bank. Never sees your real desktop. The trade you're making is "let it operate a copy of a browser logged into the apps I tell it to log into." Not "let it operate my actual machine."
|
Where you'll hear it: Anthropic shipped it first with Claude 3.5 Sonnet in late 2024. OpenAI brought it to Operator and now to GPT-5.5. Google's Gemini agents use a variant. The shared scoreboard is **OSWorld-Verified** — the benchmark that ranks how well models can drive an actual desktop. Every model release announcement now cites its OSWorld score. Now you'll know what they're bragging about.
AI doing screen-based work, real workflow →
|
|
03 |
Deep Insight
I gave AI my email inbox for a week. Here's the prompt.
|
Most AI newsletters tell you what AI can do. This is what I actually did with it for seven days, the prompt I used, and the parts that worked vs. the parts I had to babysit.
I'm not going to tell you AI fixed email. It didn't. What it did was take the sorting off my plate, leaving only the deciding — which is the part of email I'm actually paid to do. The verdict isn't "magical." It's "useful enough that I haven't turned it off after fourteen days." For me, that's the bar.
I used GPT-5.5 with the new computer-use mode and the Gmail connector. Claude Sonnet 4.6 with its Gmail integration would work the same way. The prompt below is what I pinned as a "memory" so I didn't have to re-explain it every morning. About 45 minutes of upfront writing. Roughly 25–35 minutes saved per day after that.
PROMPT Every morning at 7am, do this with my last 12 hours of inbox:
1. ARCHIVE: anything that's a newsletter, notification, or
promotion I haven't opened in 30+ days.
2. FLAG: any message from someone I've replied to in the last
90 days, anything mentioning a deadline, anything marked
URGENT in the subject.
3. DRAFT a reply (DO NOT SEND): for any clear question that
asks a yes/no decision or scheduling. Save to drafts.
4. LABEL "review": anything you're under 80% sure about. Do
not act on these. Just label them.
When done, reply in chat with one line:
"Archived X, flagged Y, drafted Z, review N." |
Four design choices are doing the work here, and getting any of them wrong turns this from useful to dangerous.
| i. | "Pin the rules in a memory." Write the rules ONCE in detail, then say "use these rules every morning." GPT-5.5 retains preferences across sessions. Claude does too. Without this, you re-explain every day and the rules drift. With it, the rules are locked. | | ii. | "Draft, never send." Critical. Computer-use agents will happily click "send" if you let them. Don't let them. Drafts give you a 5-second sanity check before a message goes out under your name. The AI is good enough that I sent maybe 80% of its drafts without major edits — but that 20% would have been horrible if it had auto-sent. | | iii. | "When in doubt, label and leave." AI's failure mode is overconfidence. Tell it: if a message doesn't clearly match a rule, do nothing and add it to a 'review' label. Better to miss something than to archive a wedding invitation as spam. After two weeks of refinement my false-positive rate dropped from ~8% to under 2%, but only because the "review" label caught the borderline cases. | | iv. | "Review the audit log on Sundays." Every computer-use agent worth using produces a weekly summary of what it touched. Read it. Look for the pattern in what you flagged as wrong. After three Sundays I'd refined the rules four times and the prompt felt like an extension of how I think about email. |
The drafts I sent (after a glance) averaged maybe 80% of the way to my voice. I edited every single one. But editing 8 drafts in 4 minutes beats writing 8 from scratch in 25 minutes. That math is the whole point. Try this on your own inbox this week. Even if you only get the rules half right, you'll learn what your inbox actually contains, which is its own gift. Reply and tell me what tripped your version up — I'll print the funniest failures in Issue 003.
|
04 |
The Workflow
One way to use AI at your job this week
|
| |
Stop reading email. Start reviewing what AI already triaged.
This is the same prompt from Section 03, with the framing that matters most: the bottleneck on email isn't reading — it's deciding what to ignore. AI is bad at writing your replies. It's great at sorting. Use it for the second job, not the first. Set the prompt up once with whatever AI client has computer-use or a Gmail connector (GPT-5.5 in ChatGPT, Claude Sonnet 4.6 with Gmail integration, or any agent platform that can read Gmail). Run it manually for three days before you let it run unattended.
Why it works: You already triage email in your head, badly, all day. You spend 30 seconds deciding whether to open something, then 20 seconds deciding whether it deserves a reply, then 90 seconds writing the reply, then 30 seconds double-guessing the reply. AI does the first two minutes for you with no emotional attachment to anyone in your inbox. It's the difference between cooking a meal and putting a delivered meal on a plate. You still eat. You don't shop, chop, or stand at a stove.
|
Do this week
Pick one weekday. Set up the prompt before bed. Wake up to a triaged inbox. If you hate the result, you've spent 45 minutes learning what triage rules actually feel like. Worth it.
|
|
|
05 |
The Side Play
One way to make money with AI this week
|
| |
| ◆ Income Idea · Play №002 |
Sell a one-week "AI assistant setup" to small business owners.
Pick a small-business niche you understand well — dentists, gyms, e-commerce stores in one category, real-estate agents in your area, freelancers in a specific trade. Productize a single offer: "I'll set up a working AI assistant for your business in five days. One use case. Done." Pick one concrete use case per niche: AI customer-service that answers your top 30 questions. AI voicemail screener with text summaries. AI lead-qualifier that runs on your website chat. Price it as a flat fee, $499–$1,500, depending on integrations. Sell on cold-email LinkedIn DMs to your existing network and niche Slack groups. Not on Upwork — that's a race to the bottom you can't win.
Why it works: Small-business owners hear "AI" and feel like they're missing out, but they don't have time to figure it out. They want someone to just do the thing. The work isn't technical — it's translation. You write the prompts. You connect the tools. They get a working result. Realistic revenue: $2,000–$6,000/month with one to two clients a week. Not $50k. Not $10k. The honest number, that's the number you can hit.
|
Do this week
Pick the niche. Write a 4-line cold pitch. Send it to ten people you already know in that niche. Don't build anything yet — sell first. Build only after someone says yes.
|
|
|
06 |
The Stack
Three tools I'm testing this week
|
Not affiliate picks. Not sponsored. Just what's open in my browser tabs.
01 |
Manus
General-purpose agent platform · Free preview
The viral one. Type a task, watch a sandboxed browser do it. Two weeks of testing — flashy demos, real cracks under pressure (failed at booking a flight twice, succeeded at scraping LinkedIn into a spreadsheet). Worth one evening to feel where computer-use agents actually are. You'll learn more from one failure than from ten demo videos.
|
02 |
Lovable
Describe-an-app builder · $20/mo
Closest thing to "non-engineer ships an app." I described a tiny CRM in three sentences and had a working version in 90 seconds. Not production-grade. Production-grade-adjacent. For internal tools and prototypes that don't need to scale, faster than anything else right now.
|
03 |
Atlas Browser
AI-native browser · Free
Replaces Chrome for the workflows where you'd otherwise paste URLs into ChatGPT. The sidebar reads any page, the agent mode browses for you, history is searchable like a notebook. Switch costs are real — extensions don't all port — but if you're an AI-heavy researcher, the speedup is significant.
|
|
07 |
Inside FindSkill
What's new for members this week
|
| New |
AI for Bedside Nurses: Smarter Charting
For nurses who want EHR charting that doesn't eat their shift. Real workflows, no compliance hand-waving. Start the course →
|
|
| New |
AI for HR: Legal-Safe Workflows
For HR teams who want AI on resumes, references, and policy drafts without a legal letter on Monday. Start the course →
|
|
| Early Access |
Course completion redesign
Reaching 70% gets you the certificate. Score is gone. Retake flow added. Instant celebration screen on completion. Pro members see it first.
|
|
|
|
What's one workflow you'd hand to an agent this week if it was good enough?
Reply to this email. I read every one. Bonus: tell me which of the three model releases you actually tried, and I'll factor it into next week's brief.
↵ Hit reply
|
|
The Skill · by FindSkill.ai
|
|