Source: raw/A_NEW_Agentic_OS_Standard_for_Claude_5.5_is_here_Claude_Mods_Jev_5.5_Models.md (current version); raw/The_NEW_Agentic_OS_standard_for_Claude_5_Models_is_here_Full_Breakdown.md (predecessor)
Creator: Robo Nuggets (Jay) · Platform: YouTube
- Current: “A NEW Agentic OS Standard for Claude 5.5 is here (Claude Mods, Jev, 5.5 Models)” · URL: https://www.youtube.com/watch?v=s_1GPN-NXA0 · Duration: 38:10 · Published: 2026-10-21
- Predecessor: “The NEW Agentic OS standard for Claude 5 Models is here (Full Breakdown)” · URL: https://www.youtube.com/watch?v=8NSyI-npJCU · Duration: 21:39 · Published: 2026-08-21
Jay’s “agentic operating system” is a Claude Code workspace organized around four layers he calls ARMS: Apps, Routines, Memory, Skills, taught bottom-up (skills first, apps last) with three levels per layer. The two videos are one framework at two points in time: the August version was built for the Claude 5 models, the October version rewrites most levels for the 5.5 models, Jev and Claude Mods. Jay’s own estimate is that the dashboard is 20-30% of the value and the workspace organization underneath is the other 70%. This article records the October version as current and keeps the August levels for comparison.
Key Takeaways
- ARMS = Apps, Routines, Memory, Skills. Learn them bottom-up. Skills and memory come before routines because routines need the confidence that the agent works unsupervised.
- Skills come in three types, by job. Database skills hold reference knowledge (tone of voice, a thumbnail library). Artifact skills produce designed output, usually HTML first, then PDF or slides. Self-triggered skills run from a dashboard button through
claude -pwith no chat open. - Guard database skills against overfitting with wildcards. If Claude pitches ten titles, have eight follow the references and two ignore them. Good wildcards get promoted into the reference set. Jay keeps a tiny
/wildcardskill that defines what “wildcard” means for every other skill. - Route each skill to the cheapest model that handles it. A spelling-only
/grammarskill runs on Haiku. Jay routes prompts through Jev: he reports 4,451 routing calls in a few weeks, about 0.6 s each, 18 cents total, with about four in ten sent to Haiku. These are his numbers from his own workspace. - Make the skills you use most self-improving. His review app attaches “update the /cut and /animate skills from these comments” to the feedback he pastes back to Claude, so every review also edits the skill.
- Routines are active or passive. Active routines hand you an output to review (a daily inbox digest over Telegram). Passive routines stay silent unless they find something (a weekly brand-misuse watch). At 50 routines, only passive ones keep the review load bearable.
- Cap expensive routines and let a routine suggest new ones. Sort routines by tokens per week and ask Claude to cut the costly ones to a budget. A Friday routine mines the week’s sessions for jobs you keep doing by hand and proposes them as routines.
- Long-term memory is three layers of plain-text routing.
MEMORY.md(loaded every session) links into every department;CLAUDE.mdis a map of the workspace that routes to department files (business.md,content.md); database skills point to where live data sits. Jay runs about 420,000 files this way with no vector embeddings. - Short-term memory is the context window and the prompt cache. Tell
/compactwhat to keep. Watch for a cold cache: after about an hour idle, the next message re-reads the whole chat at full price. Jay shows a 600K-token Opus session that would cost about $5 to re-cache. - Connect apps in a fixed order: official connector, then CLI, then API, then community MCP. CLIs return short answers; API responses are long; MCP servers load their tool list into every session whether used or not.
- Mods and Chrome extensions change the tools you already have open. A usage-bar mod shows 5-hour and 7-day limits plus cache warmth. Chrome extensions grab a page component’s code or save YouTube titles and thumbnails into a database skill, and neither costs tokens.
- Move recurring big jobs out of chat into micro-apps. Video review, carousel review and a generations gallery each became a small localhost app. Keep a reusable UI component library so Claude does not rebuild buttons and cards every time.
How the framework changed between August and October
| Layer | August 2026 (Claude 5 era) | October 2026 (Claude 5.5 era, current) |
|---|---|---|
| Skills | L1 use Anthropic’s prebuilt skills and skill-creator. L2 enrich thin skills with reference files (SKILL.md as router). L3 trigger skills headlessly with claude -p. | L1 know your skill types (database / artifact / self-triggered). L2 route skills to the right model (Haiku, Jev routing, usage dashboard). L3 make skills self-improving. |
| Memory | L1 a workspace of files (Jay found ~60,000 in his). L2 router files: CLAUDE.md plus one router per department. L3 a visual second-brain graph. | L1 long-term memory (MEMORY.md + CLAUDE.md map + database skills). L2 short-term memory (context window, compaction, prompt cache). L3 self-evolving memory (/note, /calibrate, suggested fixes). |
| Routines | L1 Claude Desktop routines. L2 always-on routines on a Hermes agent with its own cloud computer, synced through Syncthing. L3 Claude Code on a VPS, which he expected Anthropic to offer “in the future”. | L1 active vs passive routines. L2 capped routines. L3 a routine that suggests routines. |
| Apps | L1 connectors from the Desktop app. L2 a /search-connectors skill. L3 build your own connectors (CLI Printing Press) and micro-apps. | L1 connect apps, including unofficial ones, in a token-cost order. L2 mods and Chrome extensions. L3 micro-apps plus a reusable component library. |
The August Routines levels are historical. Jay described Desktop routines as running only while the computer is on and cloud-hosted Claude Code as something to expect later. The wiki’s Claude Code Routines article documents Anthropic’s cloud routines from April 2026, so his “coming soon” level already existed as a product when he recorded; his Hermes setup is a choice, not the only cloud option. The October video drops the cloud-hosting discussion.
Skills
Database skills. A skill can hold knowledge instead of steps. Jay’s tone-of-voice skill carries examples and principles for his emails and community posts; his /thumb skill indexes file paths to every shipped thumbnail plus inspiration images, with an HTML artifact for browsing them visually. Rule of thumb: if you produce or reference the same output type every week, give it a database skill.
Overfitting and wildcards. References make Claude stick to the references. Jay’s fix is a default split: most suggestions follow the database skill, a few are wildcards that ignore it. He uses this for titles, thumbnails and video animation styles, then adds good wildcards back as new references.
Artifact skills. Jay’s /robo skill is his design system. It routes to an HTML brand book (colors, fonts, components), and every PDF guide, deck and infographic is produced as HTML first and exported afterwards. He calls the result consistent where unbranded output looks “vibe-coded”. He also has Claude drop every artifact into a searchable artifacts inbox on his dashboard, so finding last week’s client deck costs no tokens. See The Unreasonable Effectiveness of HTML for the format argument.
Self-triggered skills. Some skills are better as a button. His /cleanup skill clears temporary files and processes Claude left behind; he runs it from a “skills deck” on his dashboard that also lets him pick model and effort per run. The mechanism is headless mode (claude -p, see CLI Reference), which sends one prompt and returns, and the skill writes a short report of what it fixed.
Model routing. Skills you forgot about may be running on Opus when Haiku would do, especially skills that are already step-by-step. Two moves: pin simple skills to Haiku 5.5, and add an automatic router. Jay’s router is Jev; his test for whether routing suits you is whether you ever notice being sent to a smaller model. He says he did not. A dashboard of most-used skills shows where the tokens go; his heaviest is /animate, which he uses to make his videos.
Self-improving skills. Jay reviews video cuts in his own app, pins comments with timestamp and screen position, copies them all, and pastes them to Claude. A toggle in the app appends an instruction to update the /cut and /animate skills from the same feedback. He describes this as a small-scale version of reinforcement learning. The general pattern: wherever you already give feedback, attach “and teach the skill” to it. Compare the self-improvement footer in Agent Skills Overview.
Routines
- Active vs passive. Active example: a daily inbox digest sent to Telegram. Passive examples: a weekly brand watch for sites misusing the brand name, a domain/TLD scout that pings only when a good name frees up, a machine-health check that flags slow apps or old files.
- Capped routines. Jay’s dashboard shows scheduled tasks on a week view, sortable by tokens per week. His example: a routine costing “around 18 million tokens or 4. He says Claude can usually do it by cutting unnecessary steps. Both figures are his illustration, not a benchmark.
- A routine that suggests routines. Every Friday a scheduled task reads back through the week’s sessions, finds jobs he keeps doing himself, and lists them as candidate routines in a suggestions app. He reviews the list before his weekly usage resets and sends the accepted ones back to Claude to implement. The same app has a memory section (see below).
Memory
Long-term memory: give Claude a map. Jay’s position is that a “second brain” is just the files in your workspace, and its value is faster or cheaper retrieval, not the visual.
MEMORY.md. Claude Code keeps it by default, one short line per remembered item, and loads it at the start of every session (see Auto Memory). Jay compares it to a library card catalog and deliberately links it to every department of his workspace because it is preloaded anyway.CLAUDE.mdas a workspace map. Instead of a rulebook, hisCLAUDE.mdroutes by department: when working on content, these files; when on personal matters, those. Each department has its own router (business.mdrouting to its subfolders). He notes Claude Code also readsAGENTS.md; per AGENTS.md, that happens only when noCLAUDE.mdis present, so a sharedAGENTS.mddoes not replace an existingCLAUDE.md.- Database skills as routers. A
/<business>-dataskill tells Claude where the up-to-date data or dashboard for that business or client lives.
He uses no vector embeddings at about 420,000 files and says he would add them only if retrieval started to hurt. He attributes to Boris Cherny the view that plain text beats vector and RAG setups inside Anthropic too; that is his paraphrase, not a sourced quote. Metadata on visual assets makes them findable by plain-text search. Compare Folder-as-Workspace Architecture and the 2026 Claude Code AIOS pattern.
Short-term memory: the context window and the cache.
- The 5.5 models have a 1M-token context window. Two costs grow with it: recall gets worse as the chat fills (context rot), and every prompt re-reads the whole chat.
- Compact with a purpose.
/compact keep the business decisions and the templates we createdinstead of letting Claude choose what to drop. Ask Claude once to write your standing “always keep” rules intoCLAUDE.md. - Watch the prompt cache. Jay says the cache goes cold after about an hour idle and the next message pays to load the whole chat again. That matches the wiki’s record that the main conversation on a Pro/Max subscription uses a 1-hour cache TTL (Token Optimization). His example of about 4/MTok input price at the 2x one-hour cache-write rate.
- Keep sessions warm. His sessions are told to ping themselves about hourly for up to four hours, stretching the one-hour window. He had Claude estimate the saving across 30 days of chats at about $437 at API prices; that is Claude’s estimate of his own usage. The Cache Tax mod in Claude Mods does the same job with a
/keep-warmcommand.
Self-evolving memory. “If you have to correct Claude on the same thing twice, your memory system probably missed something.”
/note: park a fix for a skill or app mid-job without derailing the current task. It runs on Haiku and appends to a per-skill backlog; he applies the whole backlog before usage resets./calibrate: for feedback too important to leave to automatic memory. It saves the feedback, links it to related memories, and fixes the skill it belongs to.- The weekly suggestions app also proposes memory fixes.
Apps
Connect anything, in token-cost order. His /search-connectors skill looks for, in order: the vendor’s official connector, a community CLI, the app’s API, then a community MCP server. He has used it for Adobe Premiere Pro and MyFitnessPal, neither of which has an official connector. The order is about weight: a CLI call returns a short answer; an API response is long and Claude reads all of it; an MCP server’s tool list loads into every new session even when unused. In August he recommended asking Claude to scan a community repo for safety before setting it up, and building missing connectors with CLI Printing Press.
Mods and Chrome extensions. A mod is “like a Chrome extension for Claude Code” (see Claude Mods). His usage-bar mod shows 5-hour and 7-day limits and whether the prompt cache is warm, at no token cost. He applies the same idea to the browser, where he reviews artifacts and runs his dashboard: a /chromify skill encodes how he builds extensions; one extension captures the underlying code of a page element he likes so Claude can rebuild it (better than a screenshot, he says); another saves YouTube titles and thumbnails into his database skill from a right-click. Load your own extensions from chrome://extensions with Developer mode on.
Micro-apps for big recurring jobs. Once a task is routine, ask whether chat is still the best interface. His examples: a video-review app with pinned comments, an Instagram carousel review app, and a generations gallery for images and video made through direct model connections instead of Higgsfield. Keep a reusable component library inside your design-system skill so Claude reuses components instead of spending tokens rebuilding them. He cites Stripe’s VP of design, Katie Dill, describing the same idea for Stripe staff and agents.
Try It
- Sort your skills into the three types. For each database skill, add a wildcard rule (“2 of every 10 suggestions ignore the references”).
- Run your two simplest, most-used skills on Haiku for a week (headless with
--model haiku, or through a router) and check whether you notice a quality drop. Jay does not show how he pins/grammarto Haiku. - List your routines and mark each active or passive. Convert any active routine whose output you usually skim into a passive one that only reports exceptions.
- Add a Friday routine that reads the week’s sessions and proposes three repeated manual jobs as new routines or skills.
- Rewrite
CLAUDE.mdas a routing map if it is a rulebook: one line per department pointing to that department’s router file. - Before stepping away from a long session, decide whether to compact with explicit keep-rules or keep the cache warm. Do not resume a 500K-token chat after a long break without one of the two.
- Run one unofficial app through the connector order (official, CLI, API, MCP) and record which one you picked and why.
Implementation
Tool/Service: Claude Code (terminal or desktop Code tab), optionally Jev for routing and a local dashboard.
Setup: No install beyond Claude Code. Jay’s prompts and guide are in his video description; the framework itself is folders, markdown router files and skills.
Cost: Mods and Chrome extensions that run as plain code cost no tokens. Routines and headless skills use plan usage or API credits like any other run.
Integration notes: A self-triggered skill is one headless call. Flags per CLI Reference; the exact command Jay uses is not shown in the transcript.
# run a skill without opening a chat, on a cheap model, from a dashboard button or cron
claude -p "/cleanup" --model haikuOpen Questions
- Routing numbers are self-reported. The 4,451 calls, 0.6 s and 18-cent figures, the “about four in ten to Haiku” share and the $437 cache-saving estimate all come from Jay’s own workspace and Claude’s analysis of it.
- Self-ping cost trade-off. A keep-warm ping re-reads the cached context at the cache-read rate each hour. Whether four hourly pings beat one cold re-cache depends on context size and how often he actually returns; neither video works this out.
- Plain text at 420,000 files. He says retrieval works without embeddings at that size, but shows no retrieval-cost measurement. The earlier Fable 5 Second Brain video, which uses the same four layers and department names and is probably by the same creator, reported a ~40% token saving from deterministic pre-filtering.
- The PDF guide and the review and suggestions apps are his own and pitched alongside his paid community; they were not reviewed here.
Related
- The 2026 Claude Code AIOS Pattern — the cross-practitioner convergence this framework now joins on all five dimensions
- Claude Mods — the mods layer, with Jay’s nine example mods
- Jev decision model — the router behind his model-routing level
- Skill Authoring Best Practices — Anthropic’s rules for the skills layer
- Claude Code Routines — Anthropic’s cloud routines
- Auto Memory — the
MEMORY.mdlayer he builds on - Folder-as-Workspace Architecture — the same routing-file approach from another creator
- Fable 5 Second Brain — an earlier four-layer video, likely the same creator