Fri, October 2, 2026
3-minute read
3 AI UPDATES · 2 THINGS YOU DIDN'T KNOW YOU CAN DO WITH AI · 1 WORKFLOW
👋 Good morning.
Gemini Deep Think can take gold at the International Math Olympiad. It cannot reliably read a clock. On ClockBench (180 clock faces, 720 questions) the best model got the time right 50.1% of the time against roughly 90% for humans, per Stanford's 2026 AI Index. Fine-tuning on 5,000 synthetic clocks only taught it the faces it had already seen. Capability is spiky, which is the case for telling these things exactly what their job is.
IN 3 MINUTES, GET 1% BETTER THAN EVERYONE YOU KNOW:
Rewrite Claude Code from the inside
Save the Gemini instructions you keep retyping
Direct a voice instead of rewriting the script
Make your coding agent write less code
Turn an old image folder into a montage
3 AI updates
Anthropic's Claude Code Mods, out October 1, let you rewrite Claude Code from the inside with a few lines of TypeScript. A mod is an event handler: Claude Code calls it when a tool runs, a prompt is submitted or part of the interface is drawn, and the handler can watch that event, change it, or take it over. In practice that means rewriting prompts, blocking commands before they run, or hiding secrets mid-screen-share. Mods ship inside plugins, so you install one with /plugin in the CLI or desktop app, on version 2.1.287 or later. The catch: that same reach is why you install only from sources you trust.
Do this: Read the getting-started guide, then pick one small project routine you’d want a trusted mod to handle. Review its code and access before installing.
Your coding agent now takes house rules.
One coupon worth knowing: Claude's artifact promotion runs through October 15 for Pro, Max and Team. After Claude creates or edits an artifact, the next 10 messages in that chat use 50% less of your five-hour session limit, and only the first 15 steps of each reply qualify. Weekly limits still count at the standard rate, and Claude Code is excluded.
What would you want a Claude Code mod to handle?
Gemini Skills save the instructions you keep retyping and run them from a slash command, per Google. You can stack several at once: a writing-style skill plus your brand guide, and attach plain text, PDFs or images as reference, with Drive files and Notebook coming. Personal accounts are rolling out globally now; Workspace follows in the coming weeks. Existing Gems migrate automatically: November for personal accounts, March 2027 for business, June 2027 for education.
Do this: Turn one repeated task into a skill with a clear output format and one good example, once it reaches your account.
Your copy-and-paste finger can take the morning off.
Eleven v4 landed Monday, and the headline is that it takes direction. You describe the delivery in plain language and drop inline tags like [laughs] or [said angrily in French accent], and ElevenLabs says it follows them more closely than previous models. Both models cover 90+ languages across ElevenCreative, ElevenAgents and the API, v4 Turbo answers in about 150ms for live agents, and voice cloning now runs on 10 seconds of audio. Those are the vendor's numbers, we haven't run a quality comparison.
Do this: Try one short script with two delivery directions before rewriting the words.
The narrator finally gets a director’s note.
2 things you didn't know you can do with AI
An agent skill that makes your coding agent reach for what already exists: your own code, the standard library, an installed dependency; before it writes anything new. Install with /plugin install ponytail@ponytail, then run /ponytail:ponytail-review on one overbuilt change and check every deletion it proposes. The catch: the always-on hooks want Node on your PATH; without it the skills still work, they just stop auto-activating. We haven't installed it.
A gallery of Opus 5.5 video work that shows the original post, the prompt behind it and a side-by-side remake, each linked to a video skill you can install, such as CreatorMotion's Title Scene Takeover. Pick one scene and one free skill for a small trial. The catch: some skills are premium, generation bills to your own provider account, and these are Skillry's suggested skills, not proof of what the original creators used.
1 workflow worth stealing
Tatiana Tsiguleva shared a Claude Opus 5.5 demo using 7,000+ of her own Midjourney images and her own Suno track. She reports about two hours of work. We haven’t reproduced it; here’s our suggested smaller starting point:
Start: Choose 20–30 owned images and an original or licensed track.
Prompt: “Plan a 30-second montage for a creator audience using these images and this track. Suggest a mood, a timed storyboard and image choices. Use 9:16. Explain your choices and wait for my approval before rendering.”
Check: Review pacing, text, aspect ratio and credits. Record the actual time and spend before scaling up.
The catch: This is our proposed prompt, not the author’s exact recipe. Use assets you own or have permission to use, and start small.
🎤 Join Kushank Live Today!
How to use Dots in ChatGPT
Build in real time and bring your questions.
📅 Friday, October 2, at 10AM Pacific / 1PM Eastern
📍HERE!
See you there!
Don't forget to rate today's post
That’s your 1% for today. Hit reply and tell me which repeated prompt you’d turn into a skill.
Kushank, AI teammates, and the human team
PracticalyAI
