Prompt Engineering

AI Topics Discussed on 03 Mar, 2026

—

by

Fraser

Creative & Visual Media

@gokayfem demonstrated the capabilities of current generative tools by creating a video using Kling V3, Nano Banana 2, ElevenLabs v3, and Claude 4.6 Opus, emphasizing how enjoyable it is to simply imagine and produce content with these models.

@levelsio praised a striking AI-generated video style, noting it as one of the first to escape the common “slop vibe” and expressing interest in seeing a full movie produced in that aesthetic.

Discussions highlighted challenges in AI-generated content detection, with Grok confidently but inaccurately identifying images and videos as AI or real, underscoring limitations in visual LLMs.

AI video has flooded meme supply chains, transforming content creation velocity.

Software Development

Guillermo Rauch (@rauchg), CEO of Vercel, reported a security incident where an agent using Claude Opus 4.6 hallucinated a nonexistent GitHub repository ID and attempted to deploy it directly via Vercel’s API, despite knowing the correct project details. The agent fabricated the ID without any prior GitHub API lookups, underscoring unique failure modes in AI-assisted engineering like inventing plausible but incorrect identifiers.

He noted the setup involved OpenClaw but attributed the issue to the agent’s tool access rather than the framework itself, reinforcing Vercel’s focus on guardrails for tools like Claude Code.

Claude Code was showcased for marketing applications, including skills and best practices, demonstrated to HubSpot executives.

Emphasis on building transferable coding skills across models like Claude, Codex, and OpenClaw, warning against unexamined marketplace downloads.

AI benchmarking focuses heavily on coding but neglects broader job tasks.

Automation & Orchestration

In the Vercel incident, powerful APIs posed risks for agentic workflows, as the model bypassed safer CLI approaches and “raw-dogged” the API after frustration, potentially enabling unintended deployments. Rauch advocated for CLI or MCP over direct API calls in agent pipelines to mitigate such hallucinations.

Discussions also touched on OpenClaw reliability issues in agentic setups.

@omarsar0 shared research testing LLM-based AI agents on Byzantine consensus games, where agents must agree on a value despite adversarial behavior. Key insight: agreement is unreliable even in benign scenarios, worsening with larger groups due to stalls and timeouts, underscoring the need for explicit consensus mechanisms in multi-agent systems.

OpenClaw dominated talks, with users cautioned against fully automating without human oversight to avoid poor outputs; it’s an amplifier requiring input for quality.

Anticipation for OpenClaw alternatives rivaling n8n for automations.

Agentic systems advanced with workable harnesses and reasoners enabling automation of day jobs, creating a “leisure class.”

Future agents negotiating autonomously, with frontier models dominating.

Queries for OpenClaw skills mimicking tools like Chronicle for slide automation.

Claude skills and workflows for marketing automation featured prominently.

Strategy & Ecosystem

@ostrisai expressed excitement over Anthropic’s new Claude for Open Source Program, which provides free Claude Max 20x access for six months to open-source maintainers and contributors, supporting the AI ecosystem.

Mentions included rollouts of OpenAI’s GPT-5.3 Instant and Google’s Gemini 3.1 Flash, noted for improved accuracy, reduced “cringe,” and cost efficiency.

Four major AI ability leaps identified: GPT-3.5, GPT-4, reasoners (o3), and agentic systems; competition tightens in the latest phase.

Upskilling in agent skills urged as highest leverage, portable across models and apps.

“Learn to market” parallels “learn to code” amid AI shifts.