Pussyfootin' A.I. kept going after the first catch-up post. Same rules as always: short clips of real things I actually do with AI, in my own voice, under a minute each, and the whole recipe given away free. The blog fell behind, so this post catches it up on the seven episodes since.
They run from a Monday jobs email to a warning for parents, then a three-part look at the setup behind everything I make, then a music library my AI can finally hear, and a one-sentence skill that explains anything like I'm five. Under each video is what it does, the recipe, and the honest catch, with more room than LinkedIn gives me.
A short list of jobs in my inbox every Monday at 5pm
Every Monday at 5pm, an email shows up with the jobs worth knowing about. One of my models scans the market against my own criteria, not a job board's algorithm, and sends me a short list. No accounts, no alerts, no doomscrolling. The market comes to me, on my terms.
How it works: it comes down to three pieces. Something that runs on a schedule, something that searches, and something that sends an email. The model is the foreman connecting all three. You write your criteria once and edit them like a doc whenever your standards change.
The recipe: paste this into Claude Code or Codex and fill in the brackets.
Set up a scheduled task that runs every Monday at 5pm. Each run: search the web for job postings from the last 7 days matching my criteria: [your role titles], [remote or your city], [anything you want to exclude]. Compile the best 5 to 10 into one short email: title, company, location, pay if listed, link, and one line on why it matched. Send it to [your email]. Save my criteria in a file I can edit like a doc.
Run it once to test, confirm the email actually lands, then let it run.
The honest catch: the criteria file is the real product. The schedule and the search are plumbing, and the list is only as good as how well you describe what you want. Also, web results can go stale between the weekly run and your click, so treat the list as leads to verify, not gospel.
A warning for parents about AI with no rules
This one is a warning, not a recipe. There is a version of AI your kid can download tonight that has no rules in it at all. No account, no age check, and it runs on the same PC they game on.
I run a local model like this on purpose, for creative work the mainstream tools refuse. That is how I know what else it will answer. I asked a normal model something genuinely dangerous and it refused. I asked the one with no rules the same thing and it just started answering. You will not see that part.
Why a filter won't catch it: once one of these models is downloaded, it runs entirely on the PC, offline. Router-level and network parental filters never see it working, because there is no traffic to catch.
What to check tonight: these are called uncensored or abliterated models, and the free apps that run them take no account or age check to install. Start by looking through the installed apps on the PC for a local model runner. Ollama and LM Studio are the common ones. If Ollama is there, open a terminal and run ollama list, which prints every model on the machine. Names containing "uncensored" or "abliterated" are the ones with the safety training stripped out. In LM Studio, the downloaded models show in the My Models list, with the same naming tells.
The honest note: plenty of people use these models for legitimate work, so finding one is not proof of anything bad. The move is not panic or a lecture. It is knowing it is there and deciding together what belongs on that machine. Most parents have no idea this category exists at all, which is the whole reason I made this.
Three computers from one laptop
The next three episodes were a set: the machines, the models, and the loop they run while I sleep. First, the machines.
I control three computers from one laptop. A Mac mini that never sleeps and runs everything. A render PC that does the heavy lifting. A network drive that holds the files from every machine in one place. One app on my MacBook turns it into any of them. No cloud, no server, just a local setup wired into one room, and I can reach that room from anywhere with an internet connection.
The three machines: the Mac mini is the hub. It runs my automations and AI sessions around the clock. The render PC handles 3D renders and local AI models, so the laptop never has to. The network drive is one shared folder every machine can read and write, so nothing gets stranded on one computer.
The recipe: the glue is Jump Desktop.
Install the free Jump Desktop Connect app on each machine you want to reach, Mac or Windows. Add your Jump Desktop account as a remote access user on each one. Then open Jump Desktop on the laptop and sign in with that same account. Every machine shows up as an icon. Tap one and you are on it. It connects through firewalls on its own, so there is no router or port setup. To keep a desktop Mac reachable, go to System Settings, then Energy, and turn on "Prevent automatic sleeping when the display is off" and "Wake for network access."
Want to test the idea for free first? Chrome Remote Desktop works across Mac and PC, and Macs have Screen Sharing built in for other Macs.
The honest catch: the machines you reach have to stay on and awake, and your home upload speed decides how smooth it feels when you are away.
The free models that run on my own PC
Every render on my PC is free. Not cheap, free. The cloud tools I use charge per generation, and the models on my own machine don't. I downloaded them once, they run offline, and nobody meters them. No account, no credits, no per-generation bill.
What runs on it: one open model turns a rough 3D frame into a photoreal image. Another turns stills into video and locks the motion to my exact camera move. A small local voice model reads the script.
The recipe: here is the model list, all free to download.
FLUX.2 Klein 9B turns a rough 3D render into a photoreal image. I add an edge-lock LoRA (RefControl canny) so the walls and windows stay where I built them. LTX-2.5 (22B) turns stills into video, and a depth LoRA locks the motion to the real camera move from my 3D scene. Kokoro-82M, run in the Voicebox app, is a small text-to-speech voice that runs on the machine. The image and video models run inside ComfyUI, a free node-based app. My PC has an RTX 5070 Ti with 16 GB of video memory.
The honest catch: free means no per-generation bill. You still pay for the hardware and the power. And one disclosure: the finished house at the very end of the video got one extra photo pass in Higgsfield, a paid cloud tool. Everything before it ran locally.
The loop my PC runs while I sleep
My PC works while I sleep, and some mornings it hands me nothing, on purpose. During the day I build 3D scenes and export the raw frames. They look like a video game: flat, and obviously computer made. Overnight, the PC turns those frames into a real clip, sharpens it to 4K, and checks its own work. Anything that fails a check gets set aside instead of shipped. Nobody touches it, and it costs $0 to run.
The recipe: one script on the PC, running local models inside ComfyUI.
Export the raw frames and the camera move from the 3D scene in Unreal Engine. FLUX.2 Klein repaints the first and last key frames photoreal, with an edge-lock LoRA so the architecture can't drift. SUPIR upscales those stills. LTX-2.5 fills in the motion between them, locked to the camera move by a depth LoRA. RTX Video Super Resolution sharpens the clip to 4K. Between steps, automatic checks look at sharpness, alignment, structure, and exposure, and a shot that fails goes to a set-aside folder instead of the finished one. In the morning I read one report: what finished, what got set aside, and why.
The honest catch: the checks are strict on purpose. When I made this episode, they had set aside all three shots on my latest test scene. I would rather wake up to zero clips than to a bad one. Same disclosure as the models episode: the photoreal house at the end got its final pass in Higgsfield, a paid cloud tool, from the same raw frame. The loop itself is local and $0.
Teaching my AI to hear my music library
My AI couldn't hear a single song in my music library. When I asked Claude or ChatGPT for music, they went by filenames. A search for "piano" found 15 results, and 4 of them were logo stings. As an editor, the right song can make the cut, and I was burning way too much time digging for it.
So I had Google's Gemini listen to all 574 tracks and write down what each one sounds like: the mood, the energy, and what kind of shot it fits. Now when I ask for something slower, with piano, I get 42 matches, each with a reason.
The recipe:
The model is Gemini 2.5 Flash through Google's Gemini API, with a free key from Google AI Studio that has a daily limit. Each song gets shrunk to a small MP3 and sent with the same instructions every time: genre, mood, energy, instruments, and 2 to 4 "best for" shot tags written for my kind of work (real estate, lifestyle). Tailor that part to your niche and the tags get much more useful. Every answer becomes one row in a spreadsheet, and Claude searches that spreadsheet as text. Tempo, key, and loudness are measured separately by a free script on my Mac.
The honest catch: Gemini guesses a tempo too, but its guesses and the measured numbers disagree more often than not, so I don't trust its tempo. It isn't perfect on instruments either. It tagged a couple of ambience tracks as "piano," so I still listen before I cut. And my sound effects are still searched by filename.
When you just need the short version
When you're deep in complex work, sometimes you just need the short version. I'm a video producer and editor, not a developer. Claude runs DaVinci Resolve for me through something called an MCP server, and I couldn't have told you what that was. So I typed /eli5 and the topic. ELI5 is short for explain like I'm five. One page, big pictures, few words. Now I get it.
What it is: /eli5 is a free skill you can add to Claude, and the whole thing is one sentence. That is all a skill is: one file of instructions your AI reads before it starts.
The recipe: /eli5 is a free community plugin by Thariq Shihipar, in Anthropic's community plugin marketplace for Claude Code and Cowork. There are two ways to add it. Ask Claude "Read and install this skill as a personal skill:" and paste that link. Or, in a terminal, run claude plugin marketplace add anthropics/claude-plugins-community, then claude plugin install eli5@claude-community. Then type /eli5 and any topic.
The rule underneath it: anything you ask for twice, make it a skill. What would your first one be?
Follow along
Pussyfootin' A.I. runs on weekdays on LinkedIn, one real move per episode, with the whole recipe given away free. No newsletter, no signup. Tomorrow's episode gets its own post here.
If you want media or workflows like this built for your own product or brand, by someone who treats the finish as the point, reach out.