Read this before creating a single file under app/. It describes the folder as it actually is; the rules of the work itself live in CLAUDE.md at the root. Everything a visitor sees lives under the language segment, and it is split into exactly two permission layers — a route in neither is a question nobody asked: Every user-facing route carries the [lang] prefix. A page created outside it exists in one language only and drops out of the language…
What this node is and how each thing is done. 🔒 Not installed yet — logs/launch.json has no started? Your only job is npm run launch in this folder, with one question at most. Cut off by a timeout — run it again, it continues where it stopped. Every ===PROGRESS=== line — one line to the person in their language. ===LAUNCH_ASK=== (a computer with 5–12 GB of memory) — the only question: give the person the words above it and…
Build process. The person names a skill they heard of — in an ad, a post, a video, from a friend («I heard about replica-skill») — or gives its link: find it, judge it the catalog's way, answer with curiosity, and after their yes keep it in the node's own skills «for later» (searched first) or hand the checked copy to the element's own agent now, marked «suggested by the person». Also when they say «смотри, интересный навык, может пригодится».
Build AI chat interfaces using ai-elements components — conversations, messages, tool displays, prompt inputs, and more. Use when the user wants to build a chatbot, AI assistant UI, or any AI-powered chat interface.
Answer questions about the AI SDK and help build AI-powered features. Use when developers: (1) Ask about AI SDK functions like generateText, streamText, ToolLoopAgent, embed, or tools, (2) Want to build AI agents, chatbots, RAG systems, or text generation features, (3) Have questions about AI providers (OpenAI, Anthropic, Google, etc.), streaming, tool calling, structured output, or embeddings, (4) Use React hooks like useChat or useCompletion. Triggers on: "AI SDK", "Vercel AI SDK", "generateText", "streamText", "add AI to my app", "build an agent", "tool calling", "structured output", "useChat".
Work skill. An element of this node comes to you: it asks how the node works — the wake text «Element <id> asks about the node's architecture: «…»» — or it tells the architect something, an alarm of its clock first of all — «Element <id> tells the architect: «…»». Every wake ends with an answer to the element: done, or «the core cannot — the architect must …» with the exact action. A question: find the answer in the core's instruction by its words, read only that section, answer in at most 10 lines with npm run task -- <id> \"Answer from the core to «<question>»: …\" (the only case where you read a file of the core). A message: pass the element's words to the person where they read you now, naming the element.
Work skill. The architect wants an element of this node to check or find out something and you to bring the answer («попроси template проверить …», «узнай у auth …»), or wants you to start or close an element's agent yourself. Also when «агент элемента … сообщил, что закончил (надзор …)» wakes you after such a task.
Scan the whole project for BROKEN / REPLACEMENT characters in localized text — a control byte (e.g. 0x13) or U+FFFD or mojibake left where an accented letter belonged, which renders as a box (e.g. "Documentación" showing "Documentaci□n"). Use when the owner reports a square / wrong glyph in the UI or content, says "broken character", "weird symbol", "mojibake", "validate encoding", or after authoring multilingual content and before closing a content step. It SHIPS SILENTLY (the file still parses), so run the audit to catch it. One project-root script, nothing external needed.
The path from confirmed cases to a working prototype, walked together with the owner. Load it the moment a case is confirmed and you are about to build something from it, when the mode is `cases`, and when the owner says "let's start" / "build it" / "what's next" — and again at the start of every session of such a product, because this path outlives dozens of them. It carries what nothing else does: the honest pact that opens the work, the four decisions that must be made before the first route, how the furniture (drawer, top menu, footer) is designed rather than inherited, how long an iteration may be, and where the prototype line runs. Who may confirm a case and where the queue lives is `use-use-cases`; this file does not repeat it.
Both natures. The field of publishing — channels on social networks, blogs, boards and forums, the voices of authors, the numbers every publication brings back, the topics that rise. The facts the method of shape-the-need weighs when a person's goal involves publishing, and how the work runs once the person chose its shape. Load it together with shape-the-need, never instead of it.
Work process. What happens to you and to the element agents when the core is built — a node update, the «Deploy core» button, npm run serve:rebuild: the build stops every agent, and after its end — good, failed or cut off — the ones that were running come back by themselves, in the same conversations. Load it before you start a core build or a node update, when you wake with the line «Node: the core build has finished», and when the person asks why an agent went off during a build or whether it comes back.
Work process. The main agent's own clock (node step 558-3). Load it when the person asks to do something on a schedule, every day or week, at a time, to remind, or to automate a recurring job of the node.
Build process. Nobody in this node covers a need — search Lagrix for a whole element (f2f, simulated today), then create a new custom element with a role, put the task into its intake, wake its agent and tell the architect how to go on. Load it when the person asks for something no element of the node does, asks for a new element or service, brings a link, an SDK or a task with no clear result (a research, «look into X») worth keeping and growing rather than answering once in chat, or when an element's agent asks the core for a new element.
Create a multilingual content document (news, blog, docs, any page copy) the scalable, co-located way: one route folder per document, a full base-language file, partial per-language override files — auto-discovered (no registry), with zero language branching in page code. Use whenever you add or translate content that must work in more than one language. Self-sufficient: pure file edits, no any external service.
Build process for the person, work process for you. A Fractera node is LAUNCHED on an Ubuntu VPS of the person's own — a fresh node from their fork that lives there around the clock, by its IP and then on its domain with a certificate for every name; the person's computer changes nothing and can launch one server after another. Load it when the person wants something to work while the computer is off, asks about a VPS, a dedicated server, «запустить сервер», «переехать на сервер», a server's IP and password, the page «Активация выделенного сервера», the records to write at a registrar for the server, certificates, sign-in on a server's domain, or when `npm run server` prints a marked line you must explain.
Both natures. The look of an element (design skills tested on a test home page, the approved design, colours and fonts) and its site settings (name, search, languages, menus, cookie banner, switches, the CONFIG link): where each is changed — the core's «Build → Design» and «Build → Site settings» of the element — and how the whole design path runs from the architect's Telegram with `npm run design`. Load it when the person asks to choose or test design skills, to make or compare test pages, to approve a design, to change colours, a font, the title, the menu, «hide sign-in», «turn on the footer», sends a screenshot or a link of a site whose style they like, asks where an element's design or settings are, or asks for a page or a section of one (a hero, pricing, questions, a comparison) and where such sections live.
Work skill. An element agent waits for permission to cross its boundary — the line ⟦node · permission · <id> · HH:MM⟧ with its call and a one-time token. The owner decided on 2026-11-09 that the core's main agent, not the architect, answers it. Judge the call itself, never the element's explanation: allow what stays inside its task and the node's laws, deny with the way inside its folder; never the node's keys, another element's code, keys or data, the session journals, or a provider's rule. Answer with npm run permit -- <token> allow|deny \"<why>\"; npm run permit alone lists who waits.
Work process. Turning an element of the node on or off — the node's «off» flag with a reason (node step 567) — and the node's load (step 663): who may sleep, the load rule's proposals and alarms, waking on a call or before a calendar entry. Load it when the person asks to turn an element off or on, to stop or start an element, to free memory, asks how loaded the node is or why an element does not run or is «off» or «asleep», answers «yes» or «no» to a proposal of the node's load rule, gets its alarm about memory, or when an element's agent passes you a request to turn itself off or on.
Add a NEW language to an EXISTING site across ALL its content, and translate the pages later without blocking. Use when the owner says "add Armenian / Spanish / French to the whole site", "make the site multilingual", "translate the whole site into X", "scale this to more languages", "add a new language to all pages/sections", or "add a locale". This is the ONLY correct way to add a language to existing content. Do NOT improvise it by hand and do NOT re-compose a group — neither adds a per-page locale to existing pages, and re-composing overwrites what is there. Two scripts in this skill's own folder do the work: fan-out-site-language.mjs (fans the language out, seeded with the default language so the site is valid instantly, no translation API) and translate-content-page.mjs (the non-blocking runner — you translate the strings later). Self- sufficient: plain Node, no external service, no other agent.
How to answer the owner when he asks what this project is, how something works, whether it can do X, or where a thing lives. Load it on "explain how this works", "what can this project do", "how is it built", "can I …", "why does it do that", and whenever the answer would reach past your own tree — the panel, the auth service, the data layer, the machine itself. The point: you have THREE circles of sight, and only the first one you may answer from memory of the files you just opened. For the second you quote a contract; for the third you say "I cannot see that from here" and name who can. A confident answer about a neighbour you never opened is the one failure this skill exists to prevent.
Plain text files in a repository that tell a coding agent how the project works: commands to run, conventions to follow and things to avoid. CLAUDE.md, AGENTS.md, cursor rules and skills are the common kinds.
CLAUDE.md or AGENTS.md?
CLAUDE.md is read by Claude Code. AGENTS.md is an open format that Codex, Cursor and other agents read. Many projects keep one and point the other at it.
What is a skill?
A folder with a SKILL.md that describes one capability, such as filling PDFs or reviewing code. The agent loads it only when the task calls for it.
Can I search my own team's files too?
Your agents already can, over MCP, limited to the files you're allowed to read. Searching them from this page is coming.