Prompt Engineering

AI Topics Discussed on 18 Jan, 2026

—

by

Fraser

AI Video Mocap and Character Swapping

@levelsio demonstrated a new Mocap feature on photoai.com, integrating Nano Banana Pro, Wan 2.2 Animate, and ElevenLabs voice library to enable character swaps while preserving original video quality.

He showcased examples like transforming into an “e-girl” vlogger and his PC avatar, noting its utility for Hollywood, Netflix, or YouTube content, predicting integration into creative workflows within a year.

@cocktailpeanut highlighted LTX-2 for lipsyncing videos by modifying only mouth movements, preserving the original footage, and shared prompting techniques for better results.

Prompting as a Skill in Video Generation

@cocktailpeanut emphasized that advanced video generation with LTX-2 relies on detailed prompts due to its strong text encoder (Gemma), calling poor results a “skill issue” rather than model limitations.

He advocated “text-to-concept” prompting for chainable concept extraction over direct output generation, predicting it could rival LoRAs as input processing improves.

Coding Agents and Vibe Coding

@simonw built and tested Cursor’s vibe-coded web browser (fastrender), noting it compiles after fixes and runs on Mac, with ~3M lines of code from agent iterations rather than memorization.

@levelsio praised a YouTuber (@bridgemindai) using multiple Claude Code terminals to “vibe code” apps toward $1M revenue, calling it the most interesting shipping effort recently.

Economic Preparation for AGI

@levelsio observed friends aged 25-40 aggressively acquiring assets (stocks, ETFs, gold, silver, real estate) to be “asset heavy” before AGI arrives, amid a shift where capital compounds without labor.

LLM Agent Memory Mechanisms

@omarsar0 shared a survey on the evolution of memory mechanisms in LLM-based agents, highlighting effective techniques.

Open Source AI Agents and Tools

@EMostaque pointed to @ii_posts’ new ii-agent (github.com/Intelligent-Internet/ii-agent) as an open-source alternative to Lovable/V0/Bolt.

@cocktailpeanut promoted local AI tools like MLX-Video for I2V with prompt templates.

Grok and Personalization

@cocktailpeanut suggested Grok-powered “For You” timelines enable voice commands like “remove porn” for instant curation.

Ideas and AI Coding Era

@danshipper noted it’s the best time to be an “ideas guy” with AI agents handling execution, recommending resources on Claude Code.

Coding Agents and Iterative Problem-Solving

Simon Willison (@simonw) explained that modern coding agents do not produce single-shot outputs for large problems; instead, they decompose tasks and iterate incrementally toward solutions. He emphasized efforts by AI labs to curb direct regurgitation of training data.

Vibe Coding with Multi-Terminal Claude AI Clusters

@levelsio showcased a developer streaming “vibe coding” sessions using a cluster of six Claude Code terminals (Anthropic’s AI coding tool) to build apps targeting $1M in revenue, calling it the most compelling shipping effort he’d seen recently.

AI Data Center Developments in Alberta

A.I.Warper (@AIWarper) pointed to the Wonder Valley AI Data Centre Park project in Alberta as a potential solution amid debates on data center siting and Canada’s wasted natural gas flaring, which exceeds LLM training power needs.