A NEW Agentic OS Standard for Claude 5.5 is here (Claude Mods, Jev, 5.5 Models)

Summary

Jay presents an updated agentic operating system (OS) built mainly around Claude 5.5, while noting that the approach can be adapted to other models. The OS combines a dashboard, custom apps, scheduled tasks, files, skills, and usage monitoring. Its organizing framework is ARMS: Apps, Routines, Memory, and Skills.

Skills: Reusable Instructions and Knowledge

Skills are saved instructions an agent can invoke. Jay describes three useful types:

• Database skills store reusable references and knowledge, such as writing guidelines, examples, or thumbnail references. To avoid overfitting to old examples, he recommends requesting some “wild card” outputs that deliberately try different approaches.
• Artifact skills define consistent design rules for outputs such as PDFs, infographics, and slide decks. Jay uses HTML as an intermediate format and keeps brand rules—such as fonts and colors—in a shared design system. He also recommends an artifact inbox so completed work can be found without asking the agent to search for it.
• Self-triggered skills can be launched from an interface rather than chat, such as a cleanup task. The interface can also let users select the model and effort level.

Jay describes three ways to improve skills:

1. Choose skill types suited to the work.
2. Route simple, well-defined skills to cheaper models, such as Haiku, rather than using a larger model for everything. Jay describes using a fast decision model, called “Jev” in the subtitles, to classify prompts and route them. In his workspace, he says thousands of routing decisions cost very little and often sent work to Haiku.
3. Make skills improve from feedback. For example, comments on video edits can be used to update the editing skills that produced them.

Routines: Scheduled Agent Tasks

Routines are tasks an agent runs on a schedule. Jay distinguishes between:

• Active routines, which produce an output for review, such as a daily email digest.
• Passive routines, which stay quiet unless they find something worth flagging, such as monitoring for unauthorized use of a brand.

He recommends tracking routine costs and setting spending or usage caps on expensive tasks. He also describes a weekly routine that reviews past sessions for repetitive manual work and suggests new automations. Suggested improvements can be reviewed and selectively applied through a custom micro-app.

Memory: Making Information Retrievable

Jay’s memory approach focuses on helping the agent find workspace information quickly and cheaply. He describes three layers of long-term memory:

1. memory.md: a short index loaded at the start of a session.
2. CLAUDE.md or AGENTS.md: workspace instructions and a map to departments and folders.
3. Database skills: additional pointers to specific knowledge, clients, or reference material.

He argues that a well-organized plain-text workspace may be sufficient for many users, without adding vector databases or other retrieval infrastructure. Metadata can also help with visual assets. Jay says his own workspace contains hundreds of thousands of files, but that his text-based map has been adequate so far.

For short-term memory, he discusses the context window and the cost of long conversations. He recommends:

• Using /compact with explicit instructions about what to preserve, such as decisions or templates.
• Monitoring prompt-cache status, since an idle conversation may lose its warm cache and make the next message more expensive.
• Optionally keeping a session active with periodic self-pings. Jay reports that his own analysis estimated this could have saved about $437 at API prices; this is a workspace-specific estimate.

For self-improving memory, he describes keeping a backlog of fixes, explicitly saving important feedback, and periodically asking Claude to suggest improvements.

Apps: Integrations, Modifications, and Custom Interfaces

Jay recommends connecting Claude to the tools a user relies on, including apps without official connectors. His suggested search order is:

1. Official connector
2. CLI tool
3. API
4. Community MCP connector

He favors lightweight options such as CLIs where practical and cautions that MCPs can add context overhead because their available tools may be loaded into sessions.

He also discusses mods—small add-ons that alter Claude Code’s appearance or behavior. His usage mod displays usage limits and prompt-cache information. He extends the same idea to browser Chrome extensions, such as tools for capturing webpage design details or saving YouTube titles and thumbnails as references.

For recurring, substantial workflows, he recommends building micro-apps rather than handling everything in chat. Examples in his OS include:

• A video-review app for viewing clips and adding timestamped comments.
• An Instagram carousel review and approval app.
• An image-and-video generation gallery.
• A dashboard for routines, skills, files, and artifacts.
• A recommendation app for reviewing proposed system improvements.

He also recommends creating a reusable UI component library so custom apps stay consistent and do not need to be rebuilt from scratch.

Guide and Tutorial Takeaways

The video functions as a practical setup guide. Jay says a downloadable PDF contains the prompts and resources shown, which viewers can give to their agent and adapt. The walkthrough’s main implementation ideas are:

• Organize an agent workspace using ARMS.
• Build reusable skills and route them to appropriately sized models.
• Track and cap scheduled-task costs.
• Create a searchable workspace map using plain text.
• Manage long-chat compaction and prompt-cache use.
• Use connectors, mods, browser extensions, and custom apps to move repetitive workflows out of chat.

Main Speakers and Sources

• Jay (Jay E / RoboNuggets): The video’s presenter and source of the Agentic OS examples and recommendations.
• Anthropic / Claude: Discussed as the provider of Claude models, Claude Code, and related features.
• Boris Cherny: Cited by Jay regarding Anthropic’s use of plain-text retrieval in large workspaces.
• Katie Dill, Stripe VP of Design: Cited in connection with reusable UI component libraries.