I started a new series this week called Pussyfootin' A.I. The idea is simple: short clips of real things I actually do with AI, in my own voice, under a minute each. No hype, no ten-step funnel, no "the secret nobody tells you." Just one real move per episode, and the whole recipe given away for free in the first comment every time.
This post is the catch-up. Three episodes went out over the last few days, and I wanted the blog to have them in one place, with more room than LinkedIn gives me. So under each video below is what it does, the exact recipe, and the honest catch. If you only read the videos, you get the idea. If you read the recipes, you can go do the thing yourself today.
The one idea underneath the series
I am a video editor, not a developer. Everything I show is something a working creative can actually use, done with tools that are free or close to it. The plumbing is not the product. Taste is the product. So I have no reason to hide the plumbing, and every reason to hand it over and let it drive the conversation. That is why the full how-to lives in the first comment on every post, not behind a newsletter.
Bolting a 550-billion-parameter model onto my setup for nothing
The launch episode. I added a 550 billion parameter model to my workflow this week, and it cost me nothing. It reads documents that are too big for my main agent to swallow, hands back a summary or a second opinion, and my main agent keeps doing the real work.
What it does: when a file is too large to read directly, your agent routes it to a huge free model that has about a million tokens of context. That model digests the whole thing and returns the gist, so your main agent never chokes on the size.
The recipe:
Make a free OpenRouter account. Then tell your agent (Claude Code or Codex, either works) that when a document is too big to read directly, it should route it through the OpenRouter API to nvidia/nemotron-3-ultra-550b-a55b:free. That model has about a million tokens of context, so it digests the whole thing and hands back a summary or a second opinion, and your main agent keeps doing the real work.
The honest catch: free tiers can train on what you send them. Keep client work out of it, or add ten dollars of credit and turn on the privacy setting.
Keeping up with AI from my phone, with zero new software
The techniques worth using change every week, and I keep up without installing anything. My Mac mini has its own Apple ID, so I can text it like a person. That thread becomes an idea inbox, and once in a while an AI agent reviews the whole inbox against how I actually work.
What it does: you capture interesting techniques on your phone as you find them, and later run one short session where an agent reads the thread, opens every screenshot, and sorts each idea into fits my pipeline, pure hype, or worth a test. Nothing runs in the background. The phone captures, the machine holds the inbox, the review happens only when you ask.
The recipe:
Give your always-on Mac its own Apple ID so you can text it. When you see a promising technique, screenshot it and text it to the Mac. That thread is now your idea inbox. Later, run one five minute session where your agent (Claude Code here, Codex works too) reads the message thread right off the machine, opens every screenshot, and reviews each technique against how you actually work: fits my pipeline, pure hype, or worth a test.
The honest catch: there is no app and nothing automated. That is the point. The capture is effortless and the review is deliberate, so nothing gets acted on until you sit down and look.
Making the work review itself so I stop shipping my own mistakes
I approved a cut that had a mistake in it. I work alone, so nobody was going to catch it but me, and I missed it. Now nothing ships until one model builds and a different one reviews, and every catch gets written into a file the next session reads first.
What it does: it turns a one-time habit into a standing rule. One model does the work, a second model reviews it against exact lines or timestamps, and the lessons get saved into a file that loads automatically at the start of every future session. The process remembers the mistake so you do not have to.
The recipe: paste this into Claude Code or Codex.
From now on, before I call any piece of work finished, have a second model review it, one builds and a different one checks, with notes that cite exact lines or timestamps, not vibes. Then make it permanent: write both of these as standing rules INSIDE this project's CLAUDE.md file (AGENTS.md if you use Codex), not just something you do this once, so every future session loads them automatically, (1) read LESSONS.md before starting work, and (2) do the second-model review before calling anything done. Create LESSONS.md now, and each time a review catches a real mistake append one line: the date, the mistake, and the rule that prevents it next time.
The honest catch: the trick is where the rule lives. A chat instruction is gone by the next session. A rule written into the config file gets loaded automatically every time. The reviews are useful, but the file that remembers them is the actual system.
Follow along
Pussyfootin' A.I. runs weekday mornings on LinkedIn, one real move per episode, full recipe in the first comment, every time. A new one goes up tomorrow.
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.