Prompt Engineering

AI Topics Discussed on 04 Mar, 2026

—

by

Fraser

Creative & Visual Media

Ostris shared a tutorial on training a modern ControlNet using Flux.2 klein 4b in AI Toolkit, covering image-to-sketch and sketch-to-image generation.

Javi Lopez announced collaboration with Bytedance’s Seedance 2.0 to address video generation moderation issues and UX problems like aspect ratio changes with reference images, while emphasizing the need for character abstractions including LoRAs, multi-angle details, voice, and expressions for consistency.

Heather Cooper discussed effective prompting in image generation, noting that in a 78-word prompt, not all words contribute equally—distinguishing between building elements and ignored ones separates prompting from directing.

She also showcased a “villain arc” image generated via Grok Imagine.

Cristóbal Valenzuela advocated for AI creative worlds that prioritize niche, surprising elements one degree off from familiarity, like rare shapes, rather than vast spatial simulations.

Fofr experimented with prompts for mysterious lost footage stills and accidental thumbnails.

AIWarper shared multiple tests of Kling 3.0 Motion Control, highlighting its professional Mocap-level motion control and character consistency via attached “elements.”

https://twitter.com/i/status/2029276498960711885

They noted it feels state-of-the-art for video generation workflows like NB PRO to Kling, and expressed interest in extending short clips into stories.

Software Development

Guillermo Rauch highlighted v0 Max Fast, powered by Fast Opus, as a super-intelligent tool for frontend and design engineering, praising its quality, speed, and consistency despite cost.

Alex Volkov noted Anthropic’s upcoming “auto mode” for Claude Code, a safer alternative to “–dangerously-skip-permissions,” aligning with enterprise needs.

Simon Willison published a new chapter on anti-patterns in Agentic Engineering Patterns, warning against inflicting unreviewed AI-generated code on collaborators, such as thousand-line PRs without basic functionality checks.

Dan Shipper discussed AI’s impact on software in an interview, noting huge speed gains on greenfield projects versus only 10% in mature codebases, with a rule that AI use doesn’t earn credit but is expected for best work.

Omar Sar noted the return of “ultrathink” in Claude Code sessions to deepen agent thinking without custom prompts.

Levelsio humorously posted about having to write code manually.

@peteromallet introduced Desloppify v0.9, a harness enabling modern LLMs to autonomously refactor large-scale messy codebases (91k+ lines) to standards rivaling world-class professionals, likening it to “Claude 6” performance.

He backs this with a $1,000 bounty challenge for anyone finding poor engineering in the codebase within 48 hours.

Automation & Orchestration

Rauch promoted Vercel Skills as the new onboarding UX, demonstrated by easily adding a Slack agent skill without handling OAuth or webhooks.

Dan Shipper highlighted an AI agent “Matthew Bolton” that transformed customer discovery processes but noted failures in synthetic customer calls.

Omar Sar emphasized treating agent prompts as executable business logic with living evals as feedback loops, rather than static unit tests; shared research showing retrieval methods matter more than memory writing strategies like summarization; and showcased NeuroSkill, a proactive offline agent using BCI signals for cognitive/emotional state modeling via tool calls.

Strategy & Ecosystem

Cristóbal Valenzuela pondered the future with “Dad, what was software?” amid AI advancements.

Alex Volkov speculated on GPT 5.4 releasing soon, tying into Thursday AI hype, and referenced Silicon Valley views on AI surpassing humans.

Jonathan Fischoff commented on compute trends, posting “compute equals revenue”; noted OpenAI buying 3-4x excess memory possibly to corner supply; discussed Dario Amodei’s memo calling OpenAI-Pentagon deals “safety theater”; and praised Nvidia’s secured supply chain dominance.

Simon Willison noted resignations of Qwen’s lead researcher and others shortly after Qwen 3.5 release.