Overview

The Claude 5.5 model family, Claude mods, and the Jev decision model introduce new capabilities for agent-based work. This article describes an agentic operating system (OS) built around the ARMS framework, which organizes an agent setup into four components: Apps, Routines, Memory, and Skills.

The framework is presented from the bottom up. Each component is described in three levels, from foundational setup to advanced practice. The approach aims to reduce token consumption, lower costs, and increase output.

The Agentic Operating System Dashboard

The system uses a home dashboard that provides access to:

• Micro apps built to work with agents, including:
  • A content system that generates and displays images and videos, used in place of a separate video editing application.
  • An Instagram carousel app where Claude stores posts for review and approval.
  • A recommendation app where the agent lists system improvements. These can be applied at the end of the week when usage remains.
• Calendar and routines, including the next scheduled routines.
• Email triage, showing messages Claude flags for attention or reply.
• On-demand skills, with selectable model and effort level.
• Artifacts, which Claude saves to the dashboard on completion, allowing retrieval without spending tokens on a search request.
• Second brain view, a visual map of the workspace. It contains approximately 400,000 to 420,000 files, many of them interconnected so the agent can locate information faster.
• Monitoring views, including:
  • A timeline of files, skills, and routines last edited by the agent.
  • A weekly routines view that highlights high-token routines for optimization.
  • A connections view showing which apps and skills consume the most tokens and which go unused, so they can be retired.

The aesthetics of the dashboard are secondary. The architecture and interface provide the practical value.

The ARMS Framework

ARMS stands for:

• Apps
• Routines
• Memory
• Skills

The framework's premise is that correct configuration of these four components places a user ahead of most people who use AI agents.

---

1. Skills

Skills are saved instructions that Claude follows when called.

Level 1: Knowing Skill Types

The following three skill types appear in this workflow.

Database skills store knowledge rather than procedures.

• A tone-of-voice skill holds principles, examples, and guidelines for writing emails and community posts. Claude uses it to produce first drafts that match the author's style.
• A thumbnail skill links to file paths of shipped thumbnails and inspiration references. Because thumbnails are visual, the skill also points to an artifact that holds the latest references.
• Any output type produced or referenced repeatedly throughout the week is a candidate for a database skill.

Over-fitting is a risk with database skills. When an agent receives only reference material, it tends to remain within that material and produce fewer fresh ideas. To prevent this, database skills can include wild cards by default:

• Example: Claude proposes 10 article titles. Eight follow the reference set. Two are wild cards that do not use references.
• Strong wild-card outputs can be added to the database skill as new style or tone references.
• The same method applies to video animation styles. A tech-focused style serves as the baseline, and Claude generates wild-card styles for comparison.
• A short shortcut skill defines the wild-card instruction, so other skills can call it.

Artifact skills produce outputs that do not work well in a chat window, such as infographics, slide decks, and PDFs.

• HTML serves as the common format. Agents generate HTML reliably, and Claude exports HTML to PDF.
• A design system skill links to a brand book written in HTML. The brand book defines colors, fonts, and other guidelines.
• A brand book prevents inconsistent, generic designs. The author recommends creating one for any business.
• An artifacts inbox, built by the agent, stores every artifact Claude creates and makes it searchable. This avoids token costs from asking Claude to locate older files.

Self-triggered skills run outside the chat.

• A cleanup skill removes temporary processes and files that Claude may leave on the computer. The user runs it when the computer slows down.
• The user can select the model and effort level for each self-triggered skill.
• The Claude-P feature activates Claude by sending a prompt in the background, without opening a chat. The skill returns a report of its actions and output.

Level 2: Optimizing Skills for Cost and Time

Model routing assigns each skill to the model it needs. Many beginners run all skills on large, expensive models. Skills with a clear, defined process can run on smaller models such as Sonnet or Haiku.

• A grammar skill that checks only spelling runs on Haiku. It responds faster and at lower cost.
• Jev is a classification model that routes prompts to the appropriate model. According to the author's workspace data:
  • Jev handled 4,451 routing calls over several weeks.
  • Each routing decision took 0.6 seconds.
  • All calls together cost 18 cents.
  • About four in ten calls routed to Haiku instead of a larger model.
  • The author reports no noticeable loss of quality from lower-model routing.
• The author states that Jev is up to 200 times faster and 400 times cheaper than frontier models.

Usage inspection identifies cost drivers. A dashboard that ranks skills by usage shows which skills consume the most tokens. The most-used skill in the author's setup is an animation skill, which is token-intensive. The dashboard also reveals skills the user has forgotten, which may warrant optimization.

Level 3: Self-Improving Skills

Skills improve when feedback is recorded and applied automatically.

• The author built the Rubric Editor application for reviewing animated video clips. Users pin comments to a timestamp and location in the clip. Comments can be entered by voice transcription.
• After review, the user copies all comments and sends them to Claude.
• A toggle in the application attaches an instruction to each prompt. The instruction asks Claude to update the cut skill and the animate skill based on the comments.
• The author describes this process as a simple form of reinforcement learning.

---

2. Routines

Routines are tasks that Claude runs on a schedule.

Level 1: Active and Passive Routines

• Active routines perform a task and return an output for review. An example is a daily inbox digest that summarizes email and sends it through Telegram.
• Passive routines run in the background. They notify the user only when they find something that requires attention.
• Example of a passive routine: a weekly brand watch that searches for websites and domains using the company brand without approval. It sends an alert only when it finds a match.
• Passive routines reduce review workload. Fifty active routines would require fifty reviews each week or day.
• Other passive routine candidates:
  • Scouting domain names or top-level domains relevant to the business and alerting when a suitable one becomes available.
  • Monitoring the computer, such as clearing old files or flagging applications that slow the device.

Level 2: Capped Routines

Each scheduled task consumes plan usage or API credits. Expensive routines are easy to overlook when many are active.

• The routines tab displays scheduled tasks across the week and shows run times.
• Routines can be sorted by weekly token consumption to identify routines that use usage without adding value.
• Example: a routine that costs around 18 million tokens or $10 in API credits per week can be capped at$4 per week. The author reports that Claude can often reduce the cost by removing unnecessary steps.

Level 3: Routines That Suggest Routines

Recurring manual tasks are strong candidates for routines.

• A Friday-morning routine reviews the author's sessions and identifies tasks the author performs repeatedly. It then suggests routines in the Rubric Records micro app.
• Rubric Records is a browser-based application that runs locally on the user's computer.
• At the end of each week, before usage resets, the author reviews the suggestions. The author adds accepted suggestions to a list, then sends the finalized list to Claude to apply to the setup.
• This cycle improves the operating system with each use.

---

3. Memory

Memory is everything Claude knows about the user and their work. It also includes how quickly and cheaply Claude can retrieve that information.

As files accumulate, searching all of them each time becomes slow and expensive. Memory can be organized in three levels.

Level 1: Long-Term Memory

Long-term memory consists of all files saved in the workspace, often called a second brain system.

• The author's second brain is a folder on the computer. A visual map displays the applications connected to Claude, routines, files split by department, and skills.
• A visual map is optional. Its value lies in faster and cheaper retrieval. The author uses the visual mainly for AI training and education.
• The visual map reflects the same files and folders as the folder structure.
• The workspace contains approximately 420,000 files. Retrieval requires a system at that scale.

The system uses three layers, all built on the principle of giving Claude a map of the workspace.

Layer 1: memory.md

• Claude Code creates and maintains memory.md by default. It stores one short line for each item Claude remembers about the user.
• Claude Code loads this file at the start of every chat.
• The author links memory.md to nodes across departments, such as content and community. This gives Claude a map to every part of the workspace.
• For a small workspace, memory.md is usually sufficient.
• Analogy: a library card catalog, which lists the location of each item before a search begins.

Layer 2: Claude.md

• Claude.md is a file of instructions that Claude reads at the start of every chat.
• Conventional use: a guide to how the workspace operates.
• The author's use: a map of the entire workspace. It routes to each department. For content work, it points to content files. For personal work, it points to personal files.
• Claude Code also reads agents.md. A workspace that already contains agents.md for other AI tools, such as Codex, can use it as the central operating rules file.
• Each department can have its own routing file. For example, business.md routes to subfolders and files within the business department.
• The author does not use vector embeddings. Claude 5.5 models locate information in plain text. Vector embeddings are an option if retrieval problems appear.
• Boris Cherny, creator of Claude Code, reported that Anthropic finds plain text more effective than earlier complex setups using vector embeddings and RAG databases, even across a workspace with millions of files.

Layer 3: Database skills

• Database skills tell Claude where information is stored.
• Example: a skill named after a business or client points to up-to-date information and dashboards for that business or client.
• A skill.md file in a database skill acts as a router. Claude infers the correct location from the prompt.

The core principle is to use all three layers and make the entire second brain searchable. Plain text works well for language models. Metadata, especially for visual assets, makes search easier.

Level 2: Short-Term Memory and Context Management

Short-term memory is the context window, which holds what Claude can see in one session.

• Claude 5.5 models have a 1 million token context window. This equals approximately the first four Harry Potter books in one chat.

Two problems arise as the context window fills:

1. Context rot. Recall becomes less accurate as the chat grows. Anthropic uses the term context rot for this effect.
2. Repeated reading cost. Every prompt causes Claude to reread the full conversation. A long chat increases cost and consumes usage faster.

Compacting with a purpose

• When a chat reaches the token limit, Claude compacts the conversation by generating a summary of it.
• The user can specify what to keep. Example: /compact with instructions to keep business decisions and templates created during the chat.
• Without instructions, Claude decides which details matter and may remove needed information.
• Standing principles for compaction can be saved. Claude typically updates Claude.md to reflect these instructions.

Prompt caching

• The prompt cache keeps the conversation in short-term memory. Rereading a cached conversation costs a small fraction of the standard price.
• The cache expires after about one hour of inactivity. The next message then reloads the full conversation at standard cost.
• Example: a chat idle for 11 hours and 30 minutes would require re-caching about 600,000 tokens. With the Opus model, this costs about $5 in API credits.
• The author uses a Claude mod that displays the prompt cache status. The author keeps the cache visible alongside the 7-day and 5-hour usage limits.
• Example: a chat with a warm cache holding 186,000 tokens would not incur the full cost of rereading those tokens. Another chat idle for about nine minutes would go cold in about 51 minutes without activity.
• Self-pinging: Each session runs a directive to ping itself for four hours, approximately once per hour. This extends the one-hour cache limit to four hours.
• Analysis: Claude can analyze the last 30 days of chats to estimate savings. The author's analysis estimated that extended caching through self-pinging would have saved about $437 at API prices.
• The author has open-sourced the usage model.

Level 3: Self-Evolving Memory

Memory improves when corrections are captured. If the same correction is needed twice, the memory system has missed something.

The author uses three methods:

1. Backlog of fixes. The /note skill records a fix during a task without interrupting the work. It uses Haiku to write the fix to a file dedicated to that skill. When usage limits are about to reset, the user asks Claude to apply all listed fixes.
2. Calibration. The /calibrate skill saves feedback, links it to related memories, and updates the associated skill. Claude Code saves some memories on its own. Important feedback is recorded explicitly.
3. Memory suggestions from Rubric Records. The Rubric Records application includes a memory section. The user selects suggestions and asks Claude to apply them.

---

4. Apps

Apps are external applications that Claude can reach outside the chat, along with applications the user builds.

Level 1: Connecting Apps

• The /search-connectors skill locates connectors for applications, including those without an official connector in the Claude desktop app.
• Examples connected by the author include Adobe Premiere Pro, which has no official connector, and MyFitnessPal, which also has none.
• The skill checks sources in the following order:
  1. Official connectors from the application maker.
  2. Command-line interface (CLI) tools published by the open-source community.
  3. Application programming interfaces (APIs). Access usually requires an API key.
  4. Community Model Context Protocol (MCP) connectors.

The order reflects token weight:

• CLIs are the lightest option. Claude runs a short command and receives the requested answer.
• APIs are heavier. Responses often contain long detail that Claude does not need. Claude reads the entire response, which consumes usage with each call.
• MCP connectors are the heaviest. Claude loads the full list of capabilities for each MCP at the start of every chat, consuming context before any question is asked, even when the connector is unused.

Level 2: Mods

A mod is a small add-on, similar to a Chrome extension, that changes how Claude Code looks and works.

• Example: the author's Robo bar displays weekly and five-hour usage limits and prompt cache information.
• Mods do not consume tokens.
• The concept extends to browsers. The author uses Google Chrome to review artifacts and to access the agentic OS dashboard.
• The /chromify skill encodes best practices for building Chrome extensions.
• Example extensions built by the author:
  • Robo Capture grabs the underlying code of a selected website element. Claude reverse-engineers the element to reproduce its appearance, rather than working from a screenshot.
  • A YouTube extension saves a thumbnail, a title, or both from a right-click into the thumbnail database skill. This removes the need to leave YouTube.
• Chrome extensions run through code and do not consume tokens.
• To load a custom extension, enable developer mode in Chrome extensions and load the extension from the computer.

Level 3: Micro Applications

When a task is repeated regularly, a chat window may not be the most efficient interface. Claude can build dedicated micro applications.

• Rubric Editor: the author reviews each video cut alongside the clip. Comments are attached to the clip directly, which is faster than typing in chat.
• Carousel app: the author uses it to view and review Instagram carousels that Claude generates.
• Generations: a collection of all images and videos generated through Claude. The app connects directly to image and video models rather than using Higgsfield. The author has a separate video on setup.
• Claude can build such applications with a few prompts.

Reusable UI component library

• A component library, part of the /robo skill, provides prebuilt components. Claude selects from the library rather than rebuilding each element.
• This reduces token use and keeps the look of team applications consistent.
• Stripe's VP of design, Katie Dill, described a similar UI library for employees and agents. It produces Stripe-branded material and unifies the look of applications built by the team.

---

Summary of the ARMS Framework

Resources

The author has compiled the prompts and resources described in the workflow into a PDF guide. The guide is intended to be provided to an AI agent, which can extract the relevant items.