Congrats on getting access, Muskan. Coming from Claude, the mental model shift is mostly about how you leverage the context window. If Claude is your analytical surgeon, Gemini Pro is your heavy-lifting archivist. Here is my experience running it: **The Strengths:** 1. **Zero-Chunk Ingestion:** Dump an entire repository or hours of transcripts directly into the 1M+ context window. The retrieval is incredibly sharp. 2. **Speed & Subagents:** The speed of Gemini 3.5 Flash is unmatched. It’s the perfect engine for spawning parallel subagents/workers to handle heavy execution tasks. 3. **Video and Media:** While ChatGPT/DALL-E holds the edge in creative image generation, Gemini is strong in media—especially with Google's focus on video generation tools (like Veo) which Claude and OpenAI aren't prioritizing as much. 4. **Ecosystem & Cost:** The Google One family sharing plan is a huge win for reducing subscription costs while getting Workspace integrations and storage. **The Catch (and why you need guardrails):** Gemini has a tendency to drift or hallucinate logic in long sessions. To run it reliably in agentic workflows, you need to build strict control skills—like self-verification, scope guards, and output format linters—to keep the model focused. In short: Keep Claude for surgical coding, but use Gemini whenever you need to audit datasets, generate media, or scale out cheap parallel workers.

Fantastic demo. Running a full media pipeline on Ultra is an expensive test of reasoning. The key to keeping it viable is reserving high-effort reasoning (Ultra) purely for planning and the final QC gate, while letting cheaper, faster models handle the actual execution steps (subtitles, commands, assets). What’s most interesting isn't just that AI made the video—it’s that the entire workflow became software. For anyone building agentic media pipelines here, how are you orchestrating the state? Are you using a centralized state machine (like a Python runner), or letting the agent write and execute the pipeline steps dynamically?

Fantastic video, Nate. The real magic isn't the single prompt itself—it's the massive coordination layer that the prompt triggers. Having Fable act as a Conductor that plans, delegates, and runs QC gates over specialized Worker agents is the only way to output product, web, and video without total state drift. The prompt is just the user interface; the multi-agent pipeline is the actual software. Models upgrade or get commoditized—but the pipeline logic remains your durable asset.

Fantastic video, Nate. "Keeping the process instead of the intelligence" is the absolute truth for sustainable agent design. Model intelligence is transient—process logic is durable. Using high-reasoning models purely as conductors and delegating the execution to cheaper, faster agents is the only way to survive API churn. Having those 5 disciplined gates written down as a model-agnostic workflow is a brilliant way to keep Sonnet/Opus grounded.

@Shiv pratap Singh Awesome. Rebuilding examples and making videos is the best way to internalize how these loops behave. Lately, I’ve been using my setup to ingest and experiment with different community workflows—like reverse-engineering n8n setups—to understand how they function and document them. The ultimate goal is to compound both knowledge and skills—for both myself and the agents operating in my system. Models and APIs change, but a structured codebase of custom skills and a documented knowledge graph is what actually compounds.

@Shiv pratap Singh Exactly, Shiv. For me, Obsidian is just the IDE—the raw Markdown files and the local directory structure are the actual database. I use Obsidian because it's a great visual editor for humans, and the wiki-link structure makes it easy for AI agents to traverse the graph. But the system is completely tool-agnostic. Because everything is stored in flat Markdown files, I can swap Obsidian for any other tool tomorrow without losing a single connection. The tools are transient, but the file schema is durable.