The product dossier — the fourth config, and the only one that decides WHAT you are allowed to build. Load it in `cases` mode before writing any code, when the owner says "add a product" / "this is for the shop" / "why is that page here", and whenever you are about to name a folder after a product. Two things here are counter-intuitive and both are load-bearing: the product's `id` means nothing and never changes, while everything you would naturally use as an identifier — the name, the address, the type — is something the owner edits freely; and what a product HAS built is never stored, only counted, because a stored list of files disagrees with the files.
The INSTALLABLE side of the site — per-language manifest, icons, iOS splash screens, the service worker and the install prompt. Load it when the owner says "make it installable", "add it to the home screen", "the icon is wrong", "it opens white when I launch it", "add offline", or when you touch `app/manifest.ts`, `app/[lang]/manifest.webmanifest`, `public/sw.js`, the register island, or the icon and splash generators. Two defects here are invisible from code: an app that cannot be installed at all because the manifest has no icons, and — the expensive one — a service worker serving yesterday's page, which also makes every measurement you take in the browser a lie.
ROLES — who sees what, and where that is actually decided. Load it before building anything behind authentication, before adding a role check, when the owner says "only managers should see this" or "let administrators do it too", and whenever a screen works for you and refuses somebody else. Two things you cannot guess: the visible gate on a page is a sign rather than a lock, and a door that fronts another service may never be more permissive than that service — get that backwards and the interface promises access, then refuses it, with the cause split across two places.
One address, or one address per record — the decision hidden inside almost every request about "a page for X", and the conversation that must happen before you build either. Load it whenever the owner asks for "a client page", "a page for each post", "a card for a user", "show one order", when he asks why there are numbers or words in an address, and — most important — whenever anything is being REMOVED from an address that already has a parameter in it. He will almost never use the words "dynamic route": that is the whole problem, and recognising the request behind his words is your job, not his. Not to be confused with `use-dynamic-pages`, which is about dynamic DATA on one fixed address; this is about the ADDRESS itself.
WHERE a route goes on disk and who may open it — the two layers, the four permission groups, and the folder shape a page must have. Load it before creating any folder under `app/`, when the owner asks for "a page for managers / for buyers / for me only", when a gate refuses a folder you just made, or when you are about to import something from a neighbouring group. The mistakes here are not stylistic: a page in the wrong layer has to be moved rather than configured, a folder one level too high turns the whole public layer into a tab and fails gates on code you never touched, and a borrowed import across permission groups breaks nothing today and everything later.
The SECTION layer — the closed catalogue of section kinds every page is built from, and what it takes to add one. Load it when a page needs a look the catalogue does not have, when you are about to write anything under `sections/`, or when the owner says "make a section that…". The decision this skill exists for is made before the first file: a new kind is a promise the platform keeps forever, and most of what feels like a new kind is a widget belonging to one route. Adding one wrongly does not fail loudly — it quietly makes the project's look un-maintainable.
The SEARCH surface of a public page — the part a visitor never sees and a search engine judges the site by. Load it whenever you create a public route, change an address, touch metadata, add a section to the sitemap, or the owner says "the page is not in Google", "the wrong text shows up in the preview", "the picture is missing when I paste the link", "add hreflang", "set up robots". Every rule here was bought with a defect that looked like nothing: the page opens, the build is green, the gate passes — and whole languages are missing from results, or the page exists for nobody because no sitemap names it. Read it BEFORE the first file: a canonical address chosen wrong is not a cosmetic mistake, it is the page asking not to be indexed.
WHEN the foreign `shadcn` skill may be used here, and what our law says louder than it. Load it before adding a shadcn component, before building a new SECTION KIND for the catalogue, before building a WIDGET, and the moment the shadcn skill tells you to compose a Dialog, pick a toast, or place markup on a page. The shadcn skill owns HOW its components compose; this one owns WHERE the result is allowed to live — and in this project there are only two places, a catalogue section and a route widget. Four of its rules collide with ours; each collision below is resolved in our favour and says why, so you can follow the foreign skill everywhere else with a clear head.
Build or change a STATIC PUBLIC page — a blog post, a footer page, the home page, any page whose content somebody authored and every visitor sees the same. Use when the owner says "add an article", "write a page about X", "add a section to that page", "translate this post", or when you are about to create any folder under app/[lang]/(publicLayer). This is the FIRST of the three page models: a folder per item, prerendered, indexed. Data that grows without limit is the second model (use-dynamic-pages) and a user's own screen is the third — do not carry these rules there. Read this before writing the first file, because the two defects this area produces are invisible until late: a page nobody can find, and a page that ships one project's identity into another's.
Build process. How work is PROVEN here — before you say «done», «it works», «ready to check». Load it when you are about to report a result, when a gate goes green and you feel finished, when something must be verified on the node, and when the owner asks for proof («show me the proof», «докажи»). The rule that changes behaviour: two proofs from DIFFERENT planes, one of them with a negative control, and a build log is never one of them — it reads the same whether the capability works or not.
The project's ready-made tools — what exists, where it lives, and why looking first is not optional. Load this BEFORE building any interface piece that feels generic: a picker, a cropper, a dialog for translations, a microphone, a code viewer, a media trimmer. Also load it when the owner says "we already made that", "why is there no window", "use the tool we built" — that sentence usually means a tool exists and was not found. Tools live in `_tools/<id>/` and are listed in `_tools/TOOLS.json`; anything built beside them is a second copy nobody asked for.
Where a translatable string is allowed to live, how many languages it owes, and who checks. Load it before adding ANY user-visible string, before creating a dictionary, when a gate says a dictionary is short of languages, when the owner says "translate this" or "why is this page English", and before promising him a translation run. Three storage forms exist and they are not interchangeable; the number of languages is decided by WHAT the thing is, not by how important it feels; and the guard only checks the dictionaries somebody registered by hand — which is how two dictionaries stood outside it for weeks while looking perfectly healthy.
Mode `cases` — the order in which work is allowed to start when the owner has switched it on: a case he confirmed, then a queue of numbered steps, and every step naming the case it serves. Load it when `PLATFORM-CONFIG.mode` is `cases`, when the owner says "add a case" / "what are we building next" / "decompose this", and before opening any step in a product that has a dossier. Two things you cannot guess: a case is confirmed ONLY by the owner and nothing in your reach can confirm one, and the decomposition step is found by its `kind`, never by matching its title — the title is a sentence somebody will reword.
Meaning-search over text you store — and the choice between it and the knowledge graph. Load it when the owner says "find by meaning", "search that isn't exact-match", "answer from our documents", when you are about to add a search box over prose, or when a search returns nothing useful and you cannot see why. The functions already exist in this project (`lib/fractera/vectors.ts`); what nobody guesses is that half of the failures here are not code — they are a document stored whole instead of in pieces, or the wrong tool chosen for the shape of the answer.
Widgets — the unique, route-owned pieces a project is recognised by. Load this BEFORE building any screen element that is not a plain content block: a table with its own behaviour, an animated showcase, a calculator-looking panel, anything a designer handed you as finished code. Also load it when the owner says "make it look like this", "here is a section from my other site", "put a widget after the hero", or asks why an animation must not sit on the page itself. Two kinds exist — `_widgets/static/<name>` and `_widgets/dynamic/<name>` — and the folder names them out loud because the two obey different disciplines.
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 skill. The person is stuck in someone else's interface — DNS records at a domain registrar, Resend, Google Cloud credentials, a form on a site, a button they cannot find in Google Docs — or asks you to do something there for them («помоги настроить DNS», «получи ключи Google», «не могу найти кнопку»). You do it in their Chrome on this computer with the Claude in Chrome tools: they sign in, you fill in and set up, irreversible steps stay theirs. Without Chrome and the extension you say what to install, or lead them by words step by step.
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.