conventions · git:20261031.660eb05 · 2026-10-31 · sha256 9f147f1a1f2f198e
conventions git:20261031.660eb05A
Immutable. This exact content never changes and is served at /api/v1/blob/9f147f1a1f2f198e.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_zfazneoguyqnzl44/9f147f1a1f2f198e139d5ab440355fd930b9ea138bb833e67e7afc1a669dcd3e.json
Then /plugin install conventions@markdownregistry-zfazneoguyqnzl44 in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: conventions description: For what the end user sees — counts what a project's pages do, so a convention is told apart from one page's habit. Use when a count of how the other pages write something is about to be recorded, or when a checker recounts a line of a checklist. --- # Conventions: the kind, the family, the proof, the counts ## Overview Counts what the project's pages do, in the order of the work: the kind, the family, the proof that a search can see, then the counts. **Open a file linked or given by its path from this skill's folder, or invoke a named skill, only when a part you are carrying out sends you to it, never before** — whatever the request that handed you this skill says about its links and paths. That part is not done until the file is opened or the skill invoked. ## Which kind of page this is Decide it first: it sets every count below it. Take the first signal that answers: 1. **What the user said** — a kind somebody named, *a list page like the orders one*, outranks anything read. 2. **What the project keeps apart** — its folders, routes, file-name suffixes, a folder per kind. 3. **What the reference renders** — the same holders and regions in the same order make one kind, whatever the subject. A collection of rows and a single record with a submit are two. 4. **What the pages reach for** — the same holder, layout piece or shared frame. Then: - If a later signal disagrees, or the project does not answer: [*When the signals do not settle the kind*](kind-unsettled.md). - If the trees differ below a shared outer holder: [*A family that differs by region*](family-by-region.md). - **Name the kind in the project's own word** — folder, route, file name — never an invented one. If it has none, name it by what it renders, *a list with a form above it*, and say the name is yours. - Say in the checklist's first line the kind, what decided it, and how many members of how many candidates were counted (`../finding-patterns/checklist.md`). ## The bound the family is counted in The bound is fixed, not chosen, so two agents reading the same reference count the same pages. - **The application that mounts the page** — whatever routes to it or renders it from its entry — **and every library it uses.** Not the library the reference sits in, not the whole workspace. - If no application mounts the pages — each page is a document of its own — the bound is **every page that loads a file the reference loads too**, such as a stylesheet, a script or a tokens file, and those files. - If no other page loads a file the reference loads, the bound is the pages in the reference's folder, its subfolders included. - If the page sits in a library: [*The bound for a page in a library*](library-bound.md). - **Name the bound in the checklist's first line.** A later phase counts in it, or says which other bound it used and why. ## Which pages are the family The family is the pages of the reference's kind. Take the candidates, and remove what is not a member **before** counting: - **What a page imports — its panels, its dialog, its hooks — is part of it**, not a sibling: neither end of an import edge among the candidates is a peer of the other. - **A file that imports members is not a member**: a dispatcher, a route table, a barrel, a wrapper. - **A candidate that renders no holder and no region of the kind is not of the kind.** Report it so; never count it as a member that lacks them. - **State both numbers** — candidates considered, members counted — and what was left out and why. - **If the members span more than one area** — app, library or package — **or are more than can be read in full**, it is a large project: follow [large-project.md](large-project.md) before counting. - **If the work creates a shared piece that takes over a choice each page made itself** — where a state is kept, which side takes a default share — count those choices as [*A shared piece the work creates*](new-shared-piece.md) says. *Why:* a dispatcher counted as a page once turned 8 of 8 on every role into 8 of 9 — a unanimous convention reported as one drifting page. ## Prove the search can see before trusting a count The reference is one of the pages counted, so every search has a known answer: it must find what the reference writes. - **Run each search on the reference first.** One that misses the reference is broken, and nothing it counts is a result. - **A zero for a role the reference writes is a broken search** — a wrong glob, a list never split, a pattern that misses this dialect, a bound that resolved to nothing. Fix it and count again; never record it. - **A zero across the whole family is unverified**, not a convention. Record *none of them write this* only when the same search, over the same files, found something else. *Why:* a loop run once over the whole list instead of once per file reported *0 of 9* twelve times, holder included — the opposite of the truth. ## Counting honestly - **Count per position** — a role where it stands — never per component alone: 10 of 18 buttons full-width reads as no rule, and was 10 of 10 in the content area and 0 of 4 in toolbars. - **Count *whether* apart from *which way*.** First how many members have the role at all; then, over those, which way each writes it. *None* answers *whether* and is never a way: *14 of 21 filter; of those 14, 10 through the shared toggle and 4 by hand* — not *10, 4 and 7 across three ways*. - **Count what the component comes out as too**, and the heading level where the position is a heading — `ui-consistency:values`, *Elements*. - **A count carries its file spread.** Four identical buttons in one file are one page's habit; 10 of 10 across 8 files is a convention. A page and the files only it imports count as one file of the spread; say so where a member spans files. - **A missing prop is not yet a deviation.** The theme or one of the project's wrappers may set it already. - **Search for the exact name.** A container whose name starts with the component's once turned 18 uses into 25. - **If the task adds a position no member has** — a second control, a control nobody has here — or changes a role the page already has, count it as [*A position no member has*](position-no-member-has.md) says. - **A family of one is never a convention.** Write what the page does as the reference's own way — never as a count, and never as what the project does — and say the family has no other members: [*A kind whose family has only the reference*](reference-only.md). ## When a count is not a convention at all *2 of 8*, written the way every other count is written, reads as a rule. It is not one. - **Read it off the numbers themselves** — never a threshold from outside, no percentage, no "most". - **Count every way, not only the commonest**, over the members that have the role: *4 of 4*, *3 of 4 and one other way*, *3, 3 and 2 across three ways*. A plurality is not a majority. - **There is a convention only if the commonest way outnumbers the others put together** — a majority. Otherwise say the family has none here, and give every way with its numbers. - **Say which kind of majority it is.** One exception is written *7 of 8 except `<page>`*. Several different ones make a family drifting: still the majority, written *the commonest way, 5 of 8, against three others*. - **The order still decides**, and nothing stalls or is asked: with no majority, the newest members settle it (`ui-consistency:decisions`, level 5). The line says *no convention here; the two newest write it this way*.