Prompt Engineering

AI Topics Discussed on 06 Mar, 2026

—

by

Fraser

Creative & Visual Media

Heather Cooper (@HBCoop_) shared multiple experiments with generative video models, including a first test of the LTX-2.3 video model using a Midjourney image, audio input, and text prompt.

She also demonstrated Midjourney to Veo 3.1 workflows for cinematic scenes, such as a tense over-shoulder shot with precise lighting and motion details.

Additional posts featured Midjourney style reference images (–sref 6395689745) for filmic visuals and a planetary coffee scene.

fofr (@fofrAI) highlighted effective prompting for Nano Banana 2, emphasizing novel angles like “You’ve never seen it from this angle before!” and generated a color photo of the Statue of Liberty in its original state.

A.I.Warper shared a video generated using LTX 2.3 in ComfyUI from audio and image inputs, rendering 481 frames at 24fps in 720×1280 resolution with ~40GB VRAM, noting impressive out-of-the-box results.

Levelsio highlighted photoai.com, a single 40k+ line PHP file generating AI photos, achieving $105k/mo revenue and $80k/mo profit.

Justine Moore (@venturetwins) highlighted the “pencilslop” movement as highly effective ragebait, featuring a video example under #BreakThePencil.

https://twitter.com/i/status/2029961922783895958

She also showcased a 10-minute short film “Punch the Monkey” created in under 48 hours for less than $100 using Seedance 2, impressing viewers who mistook it for a real animated series.

https://twitter.com/i/status/2029961922783895958

Software Development

Alex Volkov (@altryne) questioned if a “sand design” UI aesthetic is baked into Codex model weights after observing consistent outputs across different frontend tasks, crediting Teknium for the observation.

He promoted a podcast episode testing GPT 5.4 quirks post-launch, alongside discussions of Anthropic vs. Dow drama and Qwen 3.5.

Guillermo Rauch (@rauchg) reflected on the AI hype cycle, citing CLI tools breaking the internet, terminal DVD logo art, and YC CEO’s “vibe coding” addiction.

Omar Sar noted Claude Code desktop’s new local scheduled tasks feature for running regular tasks while the computer is awake, calling it huge for coding agents.

He also demoed Cursor with Kimi K2.5 building a personal HN feed in ~60 seconds, emphasizing faster models enabling rapid iteration.

Simon Willison published a new chapter on agentic manual testing, where agents manually try code to spot issues beyond automated tests.

Ethan Mollick (@emollick) compared AI tools in his Excel toolbar, noting differences across models like Claude, ChatGPT, Cowork, and Codex when generating sheets.

The Boring Marketer (@boringmarketer) recommended setting GPT-5.4 as the default model in OpenClaw, describing it as a significant intelligence upgrade over Sonnet 4.6.

Automation & Orchestration

Alex Volkov (@altryne) compared browser automation tools like doBrowser.io to Atlas, inquiring about agent modes and tool integrations such as ClawDBot vs. specialized coding agents.

Discussions touched on delegating coding tasks within agent workflows and unified hubs like Lobster Library for tracking AI generations, code, and logs.

Omar Sar highlighted scheduling and automating as key for working with coding agents via Claude Code.

Dan Shipper announced hiring a customer support specialist at Every to manage AI agents and deliver personalized experiences.

Machina (@EXM7777) shared a custom skill deploying 45 sub-agents modeled after real experts with distinct frameworks, enabling debates for comprehensive task evaluations and quality improvements.

Ethan Mollick discussed challenges in AI-to-AI communication, citing research on low discourse coherence hindering multi-agent performance despite high information density.

He praised Anthropic’s new nontechnical Cowork Skill for building agent skills, including interviews and benchmarks, marking a leap in capabilities though still needing human input.

Strategy & Ecosystem

Guillermo Rauch (@rauchg) reminisced about the current AI era’s excitement—from CLI launches dominating news to terminal animations and vibe coding—predicting fond memories from future perspectives.

Alex Volkov (@altryne) covered GPT 5.4 live launch impacts, model comparisons (e.g., Claude 4.6 vs. GPT 5.4 aesthetics), Codex quota drains, and broader trends like AI “psychosis.”

Simon Willison shared that Qwen3.5 4B outperforms GPT-4o on classic benchmarks per Claude analysis.

Omar Sar discussed Microsoft’s Phi-4-reasoning-vision-15B, a compact multimodal model for agent tasks, proving smaller models suffice for vision+reasoning.

Levelsio challenged claims of AI-driven software dev job growth, arguing charts mislead without zooming out (31% down from pre-pandemic), predicting top 10% devs with AI replace many amid Jevons paradox transition.

Machina (@EXM7777) called the $200 ChatGPT subscription the highest ROI business investment ever and questioned if GPT-5.4 represents benchmaxxing or AGI.

Ethan Mollick observed that new releases from OpenAI, Anthropic, or Google typically lead as the world’s best model until the next big drop, creating a self-fulfilling advancement cycle among the top labs.