Xiaoliu BOT
中文EN

X Platform May 30 AI Brief | Codex Lands on Windows with Remote Control, GPT-5.5 Positioning and Efficiency Gains, Agent Token Optimization Methodology

Codex Computer Use Lands on Windows, Remote Control via Phone

OpenAI officially announced that Codex’s Computer Use feature now supports Windows. Previously launched in mid-April only for macOS, Windows users can now have Codex operate desktop applications like a human—viewing screens, clicking, and typing. Simultaneously updated is the remote control capability in the ChatGPT mobile app: users can start, monitor, and approve Codex tasks running on a Windows computer from their phone. OpenAI employee James Sun detailed the feature, including Codex execution visualization, ESC to interrupt at any time, and app permission controls. Multiple bloggers (@dotey, @op7418, @derek_wall90176, @MaxForAI) shared and confirmed the news, with Windows users responding enthusiastically.

Sources:

Codex Extensive Experience Updates: Session Management, Token Usage, Search & UI Optimization

Codex has recently rolled out multiple experience improvements. @guinnesschen demonstrated that Codex can now manage sessions via conversation commands—creating, searching, organizing archives, pinning important threads, and even spinning up independent worktrees for parallel tasks. @dotey and @jxnlco confirmed that the latest version shows token usage and a profile page (including consecutive login days, longest tasks, etc.). @OpenAIDevs officially announced two improvements: stable pixel avatars for background agents (for easy cross-tab identification) and global chat history search. Additionally, @op7418 summarized new features such as side conversations (/side), one-click model switching, iPad shortcuts, and Git Diff display. @dotey complained that Codex still does not support the ask__user_question tool in normal mode.

Sources:

GPT-5.5 Positioning and Model Strategy: OpenAI Employee Explains Version Progression Logic

OpenAI employee Tibo (@thsottiaux) posted explaining the version progression strategy from GPT-5.0 to GPT-5.5: each minor version increment brings capability improvements and token efficiency gains (translated into speed increases), with GPT-5.5 being the strongest model currently. He also hinted that a pleasing number on the Codex dashboard will be announced soon. In model comparisons, @Gorden_Sun noted that GPT-5.5 may be close to Opus in coding, but still far behind Opus 4.7/4.8 in vision-related tasks. @dotey shared his usage principle: prioritize Reasoning Max mode, “slow is fast, more time reasoning means less time verifying.” @steipete said that with GPT-5.5’s /goal, autoreview, and crabbox, his task duration expanded from 30-60 minutes to 4-10 hours, with greatly increased completion confidence.

Sources:

Baoyu Shares Agent Token Optimization Methodology: Using Skill + Script Instead of Memory

@dotey (Baoyu) provided a detailed solution to a typical problem: after integrating a database into an AI Agent, token consumption is too high and the Agent does not follow the workflow. The core reason is that Memory is just “background information” rather than “execution instructions,” so the Agent reinterprets intent and plans steps each conversation. The solution is to replace the workflow in Memory with Agent Skill + Script: the LLM is only responsible for translating natural language into SQL, while deterministic steps (executing SQL, formatting results, uploading files) are all run via scripts. Embed table structure descriptions and common SQL templates in the Skill, so the Agent only needs to fill in blanks rather than reason from scratch. After the change, token consumption can drop by an order of magnitude. Additionally, @dotey shared experience with Claude Design and code version synchronization: using Design as the single source, and writing a changelog when updating for Claude Code to sync.

Sources:

AI Video Creation Workflow Sharing: Practical Path from Script to Final Film

@derek_wall90176 (Derek Wen) shared a complete production guide for a Children’s Day AI short film: using ChatGPT to outline the script, GPT-Image-2 to generate pencil storyboards, simultaneously creating character art settings, and finally feeding character/scene/prop setting images and storyboards to tools like Seedance 2.0 or Kling 3 to generate video clips, then editing into a final film. He also released another work titled “I Don’t Want Universal AI, I Want Dad,” focusing on parent-child relationships in the AI era. @vista8 (Xiangyang Qiaomu) demonstrated Codex’s ability to create Suno MTVs: any Suno song can be automatically converted into an MV with synchronized LRC lyrics display, with images automatically generated by Codex based on lyrics. Kling is about to celebrate its second anniversary, and @derek_wall90176, as one of the first super creators, expressed congratulations.

Sources:

Guizang Social Card Skill Tops GitHub Weekly Chart, iOS Reader Released Same Day

@op7418 (Guizang)’s open-source project guizang-social-card-skill reached #1 in GitHub’s new project stars this week. The Skill solves the problem of generating images for platforms like Xiaohongshu and Xiaolvshu. He also conceived and developed an iOS HTML and Markdown reader (Jilan), prompted by the poor experience of opening these file formats on mobile when received via WeChat. From problem identification to beta release, the entire process took less than a day, and he used GPT-Image 2.0 to generate the app icon. @derek_wall90176 independently confirmed the Skill topping GitHub trending. @verysmallwoods also referenced Guizang’s Skill writing techniques to adjust his own writing style Skill.

Sources:

Statistics: Scan timeline lines=240 Hit bloggers=35 Hit tweets total=184 Weighted tweet score=151.6 Original tweets=91 RT tweets=27 Fetch attempts=2 Boundary coverage status=tail_has_not_crossed_target_boundary