Record a skill

OpenAI has introduced Codex Record & Replay, allowing users to show Codex a repetitive workflow once (e.g., filling out expense reports or submitting leave requests), after which the AI converts the demonstration into an inspectable, editable skill for reuse. This feature marks a leap from AI "understanding instructions" to "imitating processes."

The essence of this feature is to record a one-time human operation into a programmable "skill," enabling AI to learn and reproduce complex workflows without explicit programming. This lowers the barrier to automating repetitive tasks but also raises new questions: when AI can imitate processes, what is the human role?

Anthropic has introduced Artifacts for Claude Code, allowing users to create functional HTML pages that can be shared with others. These pages can be used for PR reviews or real-time project dashboards. The feature is currently in beta for Team and Enterprise plans.

The launch of Artifacts extends Claude Code's output beyond code snippets to interactive web content. This partially addresses the limitations of "software thinking," but sharing and collaboration features remain restricted to enterprise users.

OpenAI is expanding its Daybreak cybersecurity program, releasing a new version of GPT-5.5-Cyber (available only to trusted partners). It claims the model can reproduce more vulnerabilities than Mythos. Additionally, OpenAI has launched Patch the Planet, a project to help maintainers fix vulnerable open-source software faster.

This move shows OpenAI's continued investment in AI security, but GPT-5.5-Cyber is limited to trusted partners, and its actual capabilities and limitations remain opaque.

Sakana AI has released the Fugu API, which selects and coordinates multiple models to handle difficult tasks. Its flagship version, Fugu Ultra, scores 73.7 on SWE-bench Pro and 82.1 on TerminalBench 2.1, roughly equivalent to the Fable level. However, sources indicate that users feel a gap in real-world usage.

The release of Fugu reflects industry exploration of multi-model collaboration, but the gap between benchmarks and actual performance reminds us that the robustness of such systems still needs verification.

Ethan Mollick pointed out in a tweet that the fundamental problem with extending Codex/Cowork/Code to all knowledge work is that they remain "software thinking"—the final result (software) is what matters, and code is the source of truth. But for much other knowledge work, the process is at least as important.

This view reveals the limitations of current AI tools when imitating human workflows: they excel at generating quantifiable outputs but struggle to capture and value tacit knowledge, judgment, and creativity in the process. When AI can "record" and reproduce workflows, humans need to rethink which parts of work truly require human involvement.