# Practical pack: KI Deep Dive of 5 October 2026 **Videos, Audio Dramas and Apps from One Prompt** Subscriptions, prompts and setups from the session, as I use them in October 2026. Prices, availability and terms change quickly; check them before you buy or connect anything. Creator claims are marked as such. Retrospective, chapters and all sources: https://zukunftbilden.eu/en/deepdive/resources/deepdive-2026-10-05 ## 01 Which subscription pays off If you choose only one tool, choose Claude Code on the 200-euro Max plan for now. For everything at full speed, I budget about 500 euros a month: 200 euros for Claude, 200 euros for Codex and the rest as API credit, for example in Google AI Studio for voices, music and video. Sol 6.1 is my reliable executor with computer use and, by my calculation, about five times cheaper than Astra. Opus 5.5 has the best taste for design. This recommendation applies to October 2026 and may change with the next model. ## 02 How I prompt Opus 5.5 Three rules from my own work: Opus should respect everything that already exists and improve it intelligently. In parallel, I let it build a second version for which it reviews every feature, rethinks the architecture and builds everything better from scratch. I keep the plan short and often skip it entirely. **Prompt · Improve what exists** ```text Review the whole project before you change anything: every feature, piece of content, data source and earlier decision. Respect everything that exists and improve it intelligently. Keep what works and justify every larger change in one sentence. Then test the most important workflows from the user’s perspective and show me what changed. ``` **Prompt · Rethink a second version** ```text Review every feature of this project and rethink the architecture. Consider how it would look and work if you built it better from scratch today. Build this second version alongside the existing one and leave the first version unchanged. Finally, compare both versions and list what the new one does better and what it still lacks. ``` ## 03 Audio drama, voices and music Create an API key in Google AI Studio and pass it to your agent through an environment variable rather than in the chat. That gives it Gemini 3.8 Flash TTS for voices with breathing, emotion and accents, and Lyria for music. Opus 5.5 writes the script, casts the voices and directs music and sound. For songs with vocals, Suno remains my first choice; commercial rights require a paid plan. If music generated in code sounds like MIDI, use a finished track from Suno or Lyria that you hold the rights to. ## 04 Images and video For image editing over many rounds, I showed FLUX 3. In the comparison, the image stayed stable after many changes, while quality with Images 2.5 declined with each round. Seedance 2.5 suits video with multiple shots and audio, and ElevenLabs’ new Eleven v4 suits voices. ## 05 Dots in Europe In the EU, Dots is currently only enabled on the Business plan, probably for data protection reasons. In the session I showed how I activated Dots once through a VPN; the free version of Proton VPN worked for this. After connecting it to Slack once, the dot remained reachable there without the VPN. Dots runs on Astra without a token limit and has its own cloud computer with Blender, GIMP, Godot, CAD tools and a terminal. Check OpenAI’s terms of use first, and whether this use is permissible for you and your data. ## 06 Computer use in Claude Code Claude Code can use the same computer use as Codex; the guide is in the linked post by argofowl. In my test, it worked in the background with its own mouse pointer while I kept working in parallel. Give the agent access only to the applications and accounts it needs for the task. ## 07 Projects, mods and Dots A project in Claude Code connects to GitHub, knows your repositories and keeps running in the cloud. While your computer is on, it fetches context from it when needed. You can close the laptop and continue from your phone. A project belongs to one topic and creates its own threads for new subtasks. Dots, by contrast, is a single agent for everything that reaches out proactively. Mods load a bundle of skills, context and plugins for a particular kind of work, such as videos, websites or presentations. My impression is that mods have not yet fully rolled out in Europe. ## 08 Jev for fast decisions Jev from TypeSafe AI returns values and probabilities. That makes many small decisions cheap: which part of a page is relevant to you, which model should answer a question or how well two profiles match. According to the provider, output is free and input is very cheap. Jev is currently in early access. ## 09 Hermes, Paperclip and local models Hermes is open source under the MIT licence and lets you choose the models freely. At first I wanted a single agent that did everything. Clearly scoped tasks and several agents work better. For that, I put Paperclip on top; it manages communication, budgets and skills. Underneath, I run Hermes, Claude Code and Codex. For local models, Can I Use LLM (caniusellm.com) and CanIRun.ai show what runs on your hardware and which quantisation you need. On arena.ai you can compare open-source models by task area. ## Links for these steps ### Models and tools from the session - [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5): Anthropic’s introduction to Opus 5.5, including use cases and pricing. - [Claude Code](https://code.claude.com/docs/en/overview): Documentation for Claude Code, the workspace where I use Opus 5.5. - [Day 1: faster default speed · Tibo](https://x.com/thsottiaux/status/2107158998495748264): According to an OpenAI employee, Astra and Sol 6.1 run about 50 percent faster at default speed. Part of the daily launch series. - [Jev · TypeSafe AI](https://typesafe.ai/blog/introducing-system-one-models-and-jev): Jev answers typed questions with values and probabilities instead of text. Currently in early access. - [System One models · documentation](https://docs.typesafe.ai/concepts/system-one): The concept behind Jev, with examples of queries and answer formats. - [FLUX 3 Image](https://bfl.ai/models/flux-3-image): Image model from Black Forest Labs. In the comparison shown, images stayed stable across many editing rounds. - [Seedance 2.5](https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5): ByteDance video model with audio, multiple shots and reference images. - [Eleven v4](https://elevenlabs.io/blog/eleven-v4): ElevenLabs’ new speech model, shown together with Seedance 2.5 during the session. - [Raphael AI](https://raphael.app/): Image generator with several models and a free tier, according to the provider. ### Audio drama, voices and music - [The Human Trial · audio drama, chapter 1](https://zukunftbilden.eu/world-modeling#hoerspiel): 21 minutes in English with voices, music and sound effects. Opus 5.5 directed the script, casting and music. - [World Modeling](https://zukunftbilden.eu/world-modeling): My books, free to read and download. - [Google AI Studio · speech generation](https://aistudio.google.com/generate-speech): Try voices directly in the browser. Requires a Google account. - [Google AI Studio · API keys](https://aistudio.google.com/api-keys): Create the key your agent uses to generate voices and music here. Treat it like a password. - [Gemini API · text-to-speech](https://ai.google.dev/gemini-api/docs/speech-generation): Documentation for Gemini 3.8 Flash TTS with style, emotion and multiple speakers. - [Gemini API · music with Lyria](https://ai.google.dev/gemini-api/docs/music-generation): Documentation for music generation with Lyria 3.5, the source of the music in the audio drama. - [Suno](https://suno.com/): For finished songs with vocals. Recommended in the session when music generated in code sounds too artificial. - [Suno · pricing and rights](https://suno.com/pricing): Commercial use requires a paid plan. Check the terms before publishing. ### Agents and computer use - [Dots · OpenAI](https://openai.com/de-DE/index/introducing-dots/): Official introduction to the always-on agent. Check EU availability and the terms of use. - [Computer use in Claude Code · argofowl](https://x.com/argofowl/status/2106389732322091475): Creator guide to using Codex computer use from Claude Code. Review permissions before setting it up. - [Claude Code mods · video](https://www.youtube.com/watch?v=LBUU8dkx8ZI): German-language video on mods: bundles of skills, context and plugins for a particular kind of work. ### Open source, multiple agents and local models - [Hermes Agent](https://github.com/NousResearch/hermes-agent): Open-source agent under the MIT licence, with a free choice of models. - [Paperclip](https://paperclip.ing/): Organises multiple agents with roles, budgets and rules for who talks to whom. - [Paperclip · GitHub](https://github.com/paperclipai/paperclip): Source code under the MIT licence. - [Can I Use LLM?](https://www.caniusellm.com/): Detects your hardware and shows which open models run locally and at which quantisation. - [CanIRun.ai](https://canirun.ai/): A second option for the same check of which open models run on your hardware. - [Arena · Leaderboard](https://arena.ai/leaderboard): Rankings by task area, including open-source models. ### Come and visit - [Dragon Dynamics](https://dragondynamics.at/): Tabletop role-playing at the Penthouse. Free to try on the first Sunday of every month. - [Colearning](https://zukunftbilden.eu/colearning): Free trial day for young people, every Saturday from 10:00 to 16:30. Get in touch beforehand. - [Freie Schule Salzburg info evening · 8 October](https://freieschulesalzburg.at/anmeldung): Thursday, 8 October 2026, 19:00 to 20:30, at Strubergasse 26 in Salzburg or online. Further info evenings on the second Thursday of each month. - [Public quest board](https://paideia-lernwelt.web.app/questboard-oeffentlich): Submit requests to the school’s children and young people, such as 3D prints or cushion covers. - [KI-Kickstart](https://zukunftbilden.eu/ki-umsetzung): In three months: a computer, team training, strategy, the first productive agents and a custom platform. 1,000 euros for a referral. - [Raustruppe · WhatsApp group](https://chat.whatsapp.com/BbshYluvrc9GCg1tUConrp): A shared experience once a month, preferably outdoors. Last time, eight of us went hiking. The complete list, including all 41 X posts shown, is on the resource page.