The Filter · Weekly AI Newsletter
AI is noisy.
I make it simple.
A weekly newsletter with AI workflows, tools actually worth your time, and tutorials for founders, creators, and AI-first operators who want results, not jargon.
Free. Every week. Unsubscribe anytime.
Every Friday · 5-minute read
Every week in The Filter
3
things worth your attention in AI
I read everything AI so you don't have to. Only what actually matters this week.
2
videos worth watching
The best stuff in AI I watched this week. No filler recommendations.
1
use case you can try today
A real workflow I built and tested. Copy it. Run it. Save hours.
Who's behind it
Hey, I'm Zach.
I'm a digital marketer who's spent the past year going deep into AI. Not just reading about it, but building with it every day. I spend 40+ hours a week inside these tools, testing what works, figuring out what doesn't, and finding the stuff that actually saves time.
This newsletter is my real thoughts, hand-written. Not AI-generated, not repurposed summaries.
Every week I write about the tools I'm trying, the workflows that are working, and the things I'd tell a friend over coffee. I'm at the forefront of this stuff, and I want to break it down so you don't have to figure it all out on your own.
My goal is that the 5 minutes you spend reading this each week is the only time you have to pay attention to AI. Tune everything else out.
There's never been a better time to be alive and learn AI.
I'm rooting for you.
Free stuff
Free guides & starter kits
Copy-paste starter kits and tutorials. Grab whatever helps.
How To Become a Master With AI In One Day
Being new to AI feels like there is too much to catch up on, and there isn't. Underneath every AI tool that exists there are only three things that change how useful it is to you: whether it knows who you are, whether it can reach the apps you work in, and whether it can remember how you like a job done. This is all three, built in one day, given away in full with every prompt. Block 1, Context: the 5 core files (AGENTS, ABOUT-ME, VOICE, ROLE, OFFER), what goes in each, an interview prompt that writes all of them for you from your real answers instead of a blank page, and the complete AGENTS file to copy as-is, which is the router that tells your AI to actually read the other four (the piece almost nobody has). Block 2, Tools: what connectors and MCPs are in plain English, the order to connect in based on hours saved per week rather than setup difficulty, and a prompt that ranks your own apps and walks you through them one at a time. Block 3, Skills: the complete check-email skill from the video given away whole, with the four-pile sort, the draft-but-never-send rule, and the ask-me-first cases, plus a skill-builder prompt that does not ask what you want to build (beginners never know) but finds the job with you, scores the candidates, writes the skill, then tests it live and fixes it until the output is right. There is a test at the end of every block so you never move on wondering whether it worked, a day-two plan, and the 5 ways people lose the day. Framework credit to Remy (AI with Remy). Written for Claude, works on any AI, no code.
Read guideThe Claude Model Cheat Sheet
The people who never hit their Claude usage limit are not using it better than you, they just know two things. This is both of them, given away in full. One, the model cheat sheet: Haiku 4.5 for high-volume mechanical work, Sonnet 5 as the default for almost everything, Opus 5 only when the task needs real judgment (it runs through your allowance around twice as fast, which is where most people quietly burn their week), and Fable 5.1 for planning complex projects and long unattended runs. Every model gets a use it for, a what it costs you, and a do not use it for, plus the 3-second rule for when you cannot tell: start with Sonnet, because you can always run it again on a stronger model but you cannot get the usage back. Two, the canary: the exact one-line prompt you paste into your global instructions so Claude calls you by name at the start of every response. The moment it stops, your conversation has gotten long enough that quality and cost are both about to slip, and it is your signal to compact or start fresh. Both versions of the line are included, along with why it works. Plus how the limits are actually built (a 5-hour session clock and a separate weekly one, shared across web, desktop, and Claude Code), and the 4 moves in order for when you hit the wall. Accurate as of September 2026.
Read guide3 Secret Claude Desktop App Features
Claude got dramatically more powerful over the last few months and its best features are the ones almost nobody opens. These are the 3 most practical ones in the desktop app, with the exact click path for each and 4 copy-paste prompts, given away in full. Live artifacts: a dashboard that goes back to your data source and pulls current numbers every time you open it, so you build the client report or subscription tracker once instead of rebuilding it monthly (includes the full dashboard prompt and the reason yours will not refresh if you built it in Chat instead of Cowork). Groups: sort your chat history into the actual areas of your work so the sidebar stops being one endless list, plus the one-line rule for when something should be a Project instead, and a prompt that has Claude design the structure with you. Scheduled tasks: the closest thing to a free employee, where you describe a job once, pick a cadence, and it runs in the cloud whether your laptop is open or not, with both setup paths, the /schedule shortcut, and two complete task prompts you can steal (a 7am morning briefing that skips newsletters and drafts the replies, and a Friday client report formatted to send). Plus the order to set them up in, and the 5 mistakes that waste your first hour, including the local-files gotcha that quietly forces a cloud task back onto your machine.
Read guide5 Free Claude MCPs That Feel Illegal to Know
Out of the box Claude cannot open a website, look up anything from today, make an image, or read the current docs for the thing you are building. These 5 MCPs fix all of it, and the whole install is one prompt you paste into Claude Code. Given away in full: the install prompt that adds all five at user scope so they work in every project, the single command for each one if you would rather go one at a time, and a copy-paste test for each so you feel it working in the first ten seconds. Perplexity for live cited research instead of guesses from old training data. Higgsfield for every top image and video model in one connection, so Claude can generate, edit, upscale, and reframe from the chat you are already in. Playwright for a real browser Claude can drive, which means it can use any tool that has no API and no MCP. Context7 for the current documentation of whatever library you are building with, which kills the broken-code-from-outdated-training bug (say the words use context7 and the guessing stops). Firecrawl for searching, scraping, crawling, and mapping any site on the internet. Plus the 4 combos almost nobody runs, where two MCPs together do something neither can alone, and one stack test prompt that puts all five on a single job. Only two need a key, both free, both about a minute.
Read guideBuild a Team of Always-On AI Agents
Grok Bot gives you named AI teammates that live on a cloud computer, sign into your actual tools, hand work to each other, and keep going after you close your laptop. No Mac mini, no gateway, no terminal. The full setup given away in full with all 5 copy-paste prompts. How to actually get access (it signs in with a Cursor account, and the SuperGrok or X Premium+ link is permanent, so get it right the first time), the install, then the brain dump into reverse prompt that makes Grok Bot design your whole bot team and then build it itself, names, descriptions, and first tasks included. The plugin shortcut: install Composio once and reach a thousand apps instead of wiring every tool by hand, plus the prompt that walks your stack one approval at a time. The five-part task structure that gets a real result instead of a summary. How to turn a task that worked into a saved skill and then a scheduled routine, in that order, with a test run before you enable it. Group chats where two to six bots hand work between themselves. And the security rules nobody mentions: all your bots share one computer, so they are not separate security walls, deleting a bot does not revoke its access, and a test run does real work. Accurate as of September 2026, early beta.
Read guideBuild an AI Phone Receptionist That Books and Transfers
Fish Audio quietly shipped a full voice agent builder: no-code console, a knowledge base you upload your own docs into, real phone numbers, warm transfers to a human, and automatic summaries after every call. This is the complete walkthrough for turning that into a phone receptionist for a real business, given away in full, every click. Clone your voice from a reel (or pick a library voice), paste the interview prompt into Claude and get the receptionist's system prompt and knowledge doc written for you, build the agent (greeting, voice, expressive mode, knowledge upload), turn on warm transfer so it briefs you on your cell before connecting the caller, set up the post-call data fields so every call gets logged with name, callback number, reason, and urgency, buy a local number and bind it, publish, then roll it out the safe way by forwarding only your missed calls. Plus the optional Step 13 with the Claude Code prompt that builds the booking endpoint so it puts appointments on your calendar itself. Real costs (Plus is $15/mo through August 31), the eight gotchas, and no Twilio, n8n, Retell, or developer required.
Read guideGo further
Want more than the guides?
I help businesses turn this stuff into real systems, and I partner with the AI tools I actually use.
The Filter
AI is noisy. Get the signal.
Every Friday. The 3 things that matter, 2 videos worth watching, and 1 workflow you can run today.
Free. Every week. Unsubscribe anytime.