Prompt Engineering

AI Topics Discussed on 08 Mar, 2026

—

by

Fraser

Creative & Visual Media

Discussions highlighted tools for generative media models, including a local database-backed explorer supporting speech, video, and image models with features like prompt remixing, JSON handling, image-to-prompt reversal, and tagging.

Experiments with altering body types and environments in generations were shared.

Javi Lopez noted Udio’s removal of download buttons for AI-generated songs, calling it nerfed functionality amid broader AI content restrictions.

Justine Moore (@venturetwins) shared examples of Seedance 2 generative videos emerging from China, emphasizing their shift toward real storytelling rather than novelty clips, with one short drama garnering over 100k likes on Rednote and calls for sequels.

Software Development

Gemini 3.1 Pro excelled in OpenClaw, demonstrating strong developer practices like frequent commits, tool usage for testing and iteration, and proactive planning.

Alex Volkov suggested testing query expansion (QMD) in OpenClaw, noting built-in capabilities for context-aware searches.

Guillermo Rauch critiqued exposed health endpoints in “vibe-coded” software and emphasized queues for content post-processing.

Jonathan Fischoff explored the concept of a programming language optimized for LLMs, noting that human-centric features like complex type systems and borrow/type inference may become less relevant.

Omar Sar highlighted the OpenDev paper, an 81-page guide on scaffolding terminal-based coding agents with dual-agent planning-execution architecture, lazy tool discovery, and adaptive context management, emphasizing the shift to CLI-native systems like Claude Code.

Goose OSS released version 1.26.0 of their open-source AI agent for automating developer tasks, featuring local inference upgrades, Telegram chat integration, and additional LLM providers.

Ethan Mollick (@emollick) discussed an Anthropic study on “vibecoding,” where over-reliance on AI for code generation hinders developers’ ability to read, write, debug, and understand code without learning gains, but using AI supportively can build skills—echoing larger RCTs in education.

Mollick also tested ChatGPT and Claude for Excel on complex 1,000-year macroeconomic data, finding both capable but ChatGPT more auditable as it stayed within Excel formulas versus Claude’s Python outputs.

Automation & Orchestration

OpenClaw was praised as an agentic harness for coding tasks, with Gemini integration enabling human-like execution and tool usage.

Workflows involved building custom tools for media exploration and prompt management.

Omar Sar discussed SkillNet, an infrastructure for creating, evaluating, and organizing over 200,000 AI agent skills with relational ontology, boosting rewards by 40% and cutting steps by 30% on benchmarks like ALFWorld.

Omar Sar also covered STRUCTUREDAGENT, a hierarchical planning system using dynamic AND/OR trees and structured memory for long-horizon web tasks, achieving 46.7% success on complex shopping benchmarks through backtracking and interpretable plans.

Machina (@EXM7777) contrasted OpenClaw’s local, customizable setup—requiring daily tweaks and troubleshooting—with Anthropic’s native features like /loop for recurring prompts, remote phone control, scheduled tasks, and persistent memory, arguing the latter delivers faster ROI for business leverage.

EXM7777 questioned the need for AI browsers given agents’ full hardware access.

Strategy & Ecosystem

Guillermo Rauch argued that coding knowledge remains crucial for effective AI prompting, system security, performance optimization, API integration, and reliability, countering claims of non-coders having an advantage; he advocated understanding data flows and endorsed xAI’s mission to elevate human intelligence.

Emerging tech like Gemini 3.1 Pro showed promise in developer tools and agentic systems.

Machina (@EXM7777) noted GPT-5.4’s improved personality makes ChatGPT conversations natural and less “cringe,” closing the gap with Claude’s engaging feel that drove users to Opus.

With universal compute access, EXM7777 asked what differentiates users now.

Ethan Mollick (@emollick) critiqued viral misinterpretations of an underpowered creativity study (n=61) showing no AI-induced drop—and sustained gains—at 30 days, urging frontier models to fact-check claims against papers.

Mollick described AI fiction writing’s illusory depth, where readers project meaning onto high-implied outputs.