`, // `

`, ``, ``, ``, ``… These are NOT custom // components: remark-mdx parses every JSX tag, lowercase or not, into the same // mdxJsxFlowElement/mdxJsxTextElement node types, so without this table they // would fall through to the generic "unknown → unwrap" rule and silently lose // their formatting (bold/italic/inline-code/heading level). // --------------------------------------------------------------------------- const HTML_HEADING_DEPTH = { h1: 1, h2: 2, h3: 3, h4: 4, h5: 5, h6: 6 }; function degradeHtmlTag(node) { const name = node.name.toLowerCase(); const children = node.children || []; if (HTML_HEADING_DEPTH[name]) { return [ { type: "heading", depth: HTML_HEADING_DEPTH[name], children: inlineOnly(children), }, ]; } if (name === "p") return [paragraph(children)]; if (name === "blockquote") return [{ type: "blockquote", children }]; if (name === "strong" || name === "b") return [{ type: "strong", children }]; if (name === "em" || name === "i") return [{ type: "emphasis", children }]; if (name === "del" || name === "s") return [{ type: "delete", children }]; if (name === "code" || name === "kbd") return [{ type: "inlineCode", value: mdastToText(children) }]; if (name === "a") { const attrs = getAttrs(node); return [{ type: "link", url: attrs.href || "", children }]; } if (name === "img") { const attrs = getAttrs(node); const src = resolveDisplayableImageSrc(attrs.src); if (src) return [{ type: "image", url: src, alt: attrs.alt || "" }]; // Relative/require()'d src can't be resolved to a real URL (see // resolveDisplayableImageSrc) — keep the alt text rather than vanish // entirely, since a bare has no children to fall back to. return attrs.alt ? [textNode(attrs.alt)] : []; } if (name === "br") return [{ type: "break" }]; if (name === "hr") return [{ type: "thematicBreak" }]; // div, span, section, article, and any other purely structural tag: unwrap. return children; } // --------------------------------------------------------------------------- // Tree walk: bottom-up so a nested custom component is already resolved // before its parent inspects `node.children`. // --------------------------------------------------------------------------- function transformChildren(children, ctx) { const out = []; for (const child of children || []) { out.push(...transformNode(child, ctx)); } return out; } function transformNode(node, ctx) { if (node.children) { node.children = transformChildren(node.children, ctx); } if (node.type === "mdxJsxFlowElement" || node.type === "mdxJsxTextElement") { return degradeComponent(node, ctx); } if (node.type === "mdxFlowExpression" || node.type === "mdxTextExpression") { // A bare {expression} in prose, outside any attribute. None exist in this // corpus outside of code fences (verified) — best-effort fallback rather // than a crash if one ever appears. return typeof node.value === "string" && node.value.trim() ? [textNode(node.value)] : []; } if (node.type === "mdxjsEsm") { return []; // import/export statements — not content. } if (node.type === "yaml") { return []; // frontmatter — the caller writes its own header instead. } if (node.type === "comment") { return []; // `` and any other raw HTML comment — not content. } if (node.type === "containerDirective" || node.type === "leafDirective") { return degradeDirective(node); } if (node.type === "textDirective") { // remark-directive's inline `:name` syntax has no word-boundary check, so // ordinary prose like "user_id:group_id", "github.com:company/x.git" or // "nginx:latest" tokenizes as a textDirective too — and every occurrence // ever found in this corpus has been exactly that, never a real directive // (this blog's own `:::tip`-style directives go through containerDirective/ // leafDirective above, not here). Not worth a per-build warning: reconstruct // the exact original characters from the source, silently. Unlike // containerDirective, a directive with no `[label]` has *no children at // all* to fall back to, so keeping `node.children` here would silently // delete the word after the colon instead. if (node.position && typeof ctx.rawSource === "string") { return [ textNode( ctx.rawSource.slice(node.position.start.offset, node.position.end.offset), ), ]; } return node.children || []; } return [node]; } function degradeDirective(node) { // Docusaurus's admonitions plugin (when available — see the try/catch at the // top of this file) annotates the directive with the resolved type/title // instead of changing its node type, so `node.type` is still "containerDirective" // here even for a fully recognised `:::tip[…]`. const type = node.data?.hProperties?.type || node.name; const title = node.data?.hProperties?.title; const label = title || ALERT_LABELS[type] || (type ? type[0].toUpperCase() + type.slice(1) : "Note"); return [ { type: "blockquote", children: [paragraph([strongText(`${label}:`)]), ...(node.children || [])], }, ]; } function degradeComponent(node, ctx) { const name = node.name; if (!name) { // React fragment <>… return node.children || []; } const isHtmlTag = name[0] === name[0].toLowerCase(); if (isHtmlTag) { return degradeHtmlTag(node); } const handler = COMPONENT_RULES[name]; if (handler) { return handler(node, ctx) || []; } // The one rule that can never be skipped: keep the content, drop the // wrapper, and tell the caller so it can warn about the gap. ctx.unknownComponents.add(name); return node.children || []; } // --------------------------------------------------------------------------- // Public entry point // --------------------------------------------------------------------------- // Parse-only processor, always format:"mdx" regardless of the file's own .md/.mdx // extension — see the createProcessor import comment above for why. Never call // `.run()`/`.compile()`/`.process()` on this one: createProcessor bakes in the // full remark→rehype→recma→JS-estree compile chain, and `.parse()` is the only // method that stops cleanly at the mdast stage this file actually wants. function buildParser() { return createProcessor({ format: "mdx" }) .use(remarkFrontmatter) .use(remarkGfm) .use(remarkComment) .use(remarkDirective); } // Separate, bare processor: `.run()`-only, used solely to apply Docusaurus's // own admonitions AST transform (`:::tip[…]` → data.hProperties) to the tree // `buildParser` produced. Kept off the parser above because a bare unified() // has no recma/JS-compile stage to accidentally trigger. const directiveTransformer = admonitionsPlugin ? unified().use(admonitionsPlugin, true) : null; const stringifier = unified().use(remarkStringify).use(remarkGfm); /** * @param {string} rawSource - the full .md/.mdx file content, frontmatter included. * @param {object} opts * @param {string} opts.currentFileDir - absolute directory of the article (for * resolving relative `source="./…"` attributes). * @param {string} [opts.projectRoot] - defaults to process.cwd(). * @returns {Promise<{ markdown: string, unknownComponents: string[] }>} */ async function mdxToMarkdown(rawSource, { currentFileDir, projectRoot = process.cwd() }) { let preprocessed = escapeMarkdownHeadingIds(rawSource); if (admonitionsPlugin) { preprocessed = admonitionTitleToDirectiveLabel(preprocessed, ADMONITION_KEYWORDS); } const tree = buildParser().parse(preprocessed); if (directiveTransformer) { await directiveTransformer.run(tree); } const ctx = { currentFileDir, projectRoot, rawSource: preprocessed, unknownComponents: new Set(), }; tree.children = transformChildren(tree.children, ctx); const markdown = stringifier.stringify(tree); return { markdown, unknownComponents: [...ctx.unknownComponents].sort() }; } module.exports = { mdxToMarkdown }; ```` ### L'orchestrateur {#the-orchestrator} Le plugin s'exécute dans `postBuild`, dégrade chaque article publié, écrit les miroirs, puis construit les deux index à partir de ce qui a réussi : **plugins/markdown-export-plugin/index.cjs** ```cjs /** * @fileoverview Docusaurus Plugin: markdown-export-plugin * * Purpose: * Writes a plain-Markdown mirror of every published blog post next to its HTML * page (`/blog/` → `/blog/.md`), plus a site-wide `/llms.txt` index * and a per-series `/llms/.txt` bundle — so a reader (human or an * LLM) can fetch the article's full content without React, JSX, or the 107 * `defaultOpen={false}` accordions this corpus collapses in the HTML. * * The actual MDX → Markdown degradation lives in ./degrade.cjs; this file only * orchestrates: find the posts that really shipped, degrade each one, write * the mirrors, then build the two kinds of index on top of what didn't fail. * * Usage: * - Add to docusaurus.config.js `plugins` array: `"./plugins/markdown-export-plugin/index.cjs"` * - Runs only in `postBuild` (i.e. only during `yarn build`, never `yarn start`), * same as every other postBuild plugin in this repo. */ const fs = require("fs"); const path = require("path"); const frontMatter = require("front-matter"); const { mdxToMarkdown } = require("./degrade.cjs"); const BLOG_DIR = "blog"; /** Recursively collects every article file under a directory (mirrors the same * pattern used by scripts/internal-link-opportunities.mjs and blog-feed-plugin). */ function findPosts(dir) { const found = []; for (const entry of fs.readdirSync(dir, { withFileTypes: true })) { const target = path.join(dir, entry.name); if (entry.isDirectory()) { found.push(...findPosts(target)); } else if (entry.name === "index.md" || entry.name === "index.mdx") { found.push(target); } } return found; } // Mirrors src/components/Blog/utils/slug.ts's createSlug() exactly. Duplicated // rather than imported: that file is ESM written for Webpack/browser bundling // (`export function`), which a plain Node `require()` here cannot load without // a bundler. Keep both in sync if the slugify algorithm ever changes. function createSlug(text) { return text .replace(/([a-z0-9])([A-Z])/g, "$1-$2") .toLowerCase() .normalize("NFD") .replace(/[\u0300-\u036f]/g, "") // strips diacritics .replace(/[^a-z0-9\s-]/g, "") .replace(/\s+/g, "-") .replace(/-+/g, "-") .replace(/^-+|-+$/g, ""); } // Mirrors src/components/Blog/utils/posts.ts's permalink derivation exactly — // same "why duplicated" rationale as createSlug above. function permalinkFor(attributes, dir) { if (attributes.slug) { return attributes.slug.startsWith("/") ? attributes.slug : `/blog/${attributes.slug.replace(/^\//, "")}`; } return `/blog/${dir}/`; } /** * @param {string} siteDir * @param {string} [translationDir] when building a non-default locale, the matching * `i18n//docusaurus-plugin-content-blog` tree. Any article that has a file there is * read FROM there, so the mirror and the llms.txt index carry the translated title, * description and body. Without it, `/fr/blog/.md` served the English Markdown under a * French URL and `/fr/llms.txt` advertised English titles — the content equivalent of the * duplicate-content problem plugins/i18n-seo-guard solves for HTML. See TODO 0119. */ function loadPosts(siteDir, translationDir = null) { const blogDir = path.join(siteDir, BLOG_DIR); const posts = []; for (const file of findPosts(blogDir)) { let sourceFile = file; if (translationDir) { const relative = path.relative(blogDir, file); const translated = path.join(translationDir, relative); if (fs.existsSync(translated)) sourceFile = translated; } const raw = fs.readFileSync(sourceFile, "utf-8"); const { attributes } = frontMatter(raw); // A Markdown file under blog/ without a title isn't an article (a // fragment, a note) — same rule posts.js applies. if (!attributes.title) continue; const dir = path.relative(blogDir, path.dirname(file)).split(path.sep).join("/"); posts.push({ file, raw, // Deliberately the ENGLISH article's folder, even when `raw` came from the translation: // `files/` and `images/` are co-located with the source, never duplicated under i18n/. currentFileDir: path.dirname(file), permalink: permalinkFor(attributes, dir), title: attributes.title, description: attributes.description || "", date: attributes.date, mainTag: attributes.mainTag || null, series: attributes.series || null, draft: attributes.draft || false, unlisted: attributes.unlisted || false, }); } return posts; } /** * Absolute URL for a site-relative path, honouring the locale's baseUrl. * * `siteConfig.url` is the bare origin and `siteConfig.baseUrl` carries the locale segment * (`/` for the default locale, `/fr/` for French). Concatenating `url + permalink` skips the * locale entirely, so the French llms.txt advertised English URLs for translated articles. * See TODO 0119. */ function absoluteUrl(siteConfig, sitePath) { const base = siteConfig.baseUrl.endsWith("/") ? siteConfig.baseUrl : `${siteConfig.baseUrl}/`; return `${siteConfig.url}${base}${String(sitePath).replace(/^\//, "")}`; } // YAML frontmatter dates (`date: 2024-02-23`) are auto-typed as JS Date // objects by the YAML parser front-matter uses — string-concatenating one // directly calls its verbose default toString(). Normalize to plain // YYYY-MM-DD instead. function formatDate(date) { if (!date) return "—"; return date instanceof Date ? date.toISOString().slice(0, 10) : String(date); } // The comment carries the same info a human reader would otherwise lose once // this file is copied or pasted elsewhere: where the rendered original lives // (images, interactive components, the lot). Kept as an HTML comment — invisible // in Markdown, so a reader pasting the file into a renderer isn't shown a stray // metadata block, while it still reads as plain text in a raw view or in // whatever an LLM is handed. // // Deliberately no build timestamp. One wall-clock value rewrote all 247 mirrors // plus the 25 llms bundles on every single build (5.2 MB), which the deploy then // had to re-upload even when no article had changed. The prose below already // says the file is generated at build time, and `published` carries the date a // reader actually needs. function buildMetadataComment(post, siteConfig) { const url = absoluteUrl(siteConfig, post.permalink); return [ "", "", ].join("\n"); } function buildHeader(post, siteConfig) { const lines = [buildMetadataComment(post, siteConfig), `# ${post.title}`, ""]; if (post.description) { lines.push(`> ${post.description}`, ""); } lines.push("---", ""); return lines.join("\n"); } function writeLlmsTxt(outDir, siteConfig, posts, seriesFiles) { const byTag = new Map(); for (const post of posts) { const tag = post.mainTag || "uncategorized"; if (!byTag.has(tag)) byTag.set(tag, []); byTag.get(tag).push(post); } const lines = [`# ${siteConfig.title}`, ""]; if (siteConfig.tagline) lines.push(`> ${siteConfig.tagline}`, ""); lines.push( `${posts.length} articles. Each has a plain-Markdown mirror at its own permalink ` + `plus \`.md\` (e.g. \`${absoluteUrl(siteConfig, "/blog/.md")}\`).`, "", ); // This is the only place the per-series bundles are ever linked from — without // it they exist on disk but are unreachable from anything an LLM would fetch. if (seriesFiles.length > 0) { lines.push( "## Series (full-text bundles)", "", "Each link is every article in that series, concatenated in reading order — " + "one request instead of many.", "", ); for (const s of seriesFiles) { lines.push( `- [${s.name}](${absoluteUrl(siteConfig, `/llms/${s.slug}.txt`)}) — ${s.count} article(s)`, ); } lines.push(""); } for (const tag of [...byTag.keys()].sort()) { lines.push(`## ${tag}`, ""); const sorted = [...byTag.get(tag)].sort((a, b) => a.title.localeCompare(b.title)); for (const post of sorted) { const desc = post.description ? ` — ${post.description}` : ""; lines.push(`- [${post.title}](${absoluteUrl(siteConfig, post.permalink)})${desc}`); } lines.push(""); } fs.writeFileSync(path.join(outDir, "llms.txt"), lines.join("\n"), "utf-8"); } // Returns the series that were actually written, sorted by name — this list is // what lets writeLlmsTxt() link to them; without it they'd be orphaned on disk. function writeSeriesFull(outDir, posts) { const bySeries = new Map(); for (const post of posts) { if (!post.series) continue; if (!bySeries.has(post.series)) bySeries.set(post.series, []); bySeries.get(post.series).push(post); } if (bySeries.size === 0) return []; const seriesDir = path.join(outDir, "llms"); fs.mkdirSync(seriesDir, { recursive: true }); const written = []; for (const [seriesName, seriesPosts] of bySeries) { const sorted = [...seriesPosts].sort((a, b) => new Date(a.date) - new Date(b.date)); const chunks = [ `# ${seriesName}`, "", `A series of ${sorted.length} article(s), in reading order.`, "", ]; for (const post of sorted) { chunks.push(post.markdown, "", "---", ""); } const slug = createSlug(seriesName); fs.writeFileSync(path.join(seriesDir, `${slug}.txt`), chunks.join("\n"), "utf-8"); written.push({ name: seriesName, slug, count: sorted.length }); } return written.sort((a, b) => a.name.localeCompare(b.name)); } module.exports = function markdownExportPlugin() { return { name: "markdown-export-plugin", async postBuild({ outDir, siteDir, siteConfig, routesPaths, i18n }) { const currentLocale = i18n?.currentLocale; const defaultLocale = i18n?.defaultLocale ?? siteConfig.i18n?.defaultLocale; const isTranslatedLocale = Boolean( currentLocale && currentLocale !== defaultLocale, ); const posts = loadPosts( siteDir, isTranslatedLocale ? path.join(siteDir, "i18n", currentLocale, "docusaurus-plugin-content-blog") : null, ); // In a non-default locale, mirror and index ONLY the articles really translated into it. // `loadPosts()` reads `blog/`, i.e. the English corpus, and Docusaurus's i18n fallback // gives every one of them a live `/fr/` route — so without this, `/fr/llms.txt` would // advertise 257 English articles as the French corpus, and `/fr/blog/.md` would // mirror English prose under a French URL. See TODO 0119. let localeFiltered = posts; if (isTranslatedLocale) { const { collectTranslations, } = require("../translations-manifest-plugin/index.cjs"); const translated = new Set(collectTranslations(siteDir)[currentLocale] ?? []); // These post objects carry a `permalink`, not a `slug` — assuming the latter silently // filtered everything out (`translated.has(undefined)` is always false). localeFiltered = posts.filter((post) => { const match = String(post.permalink ?? "").match(/blog\/([^/]+)\/?$/); return match ? translated.has(match[1]) : false; }); console.log( `[markdown-export] locale ${currentLocale}: ${localeFiltered.length} translated article(s).`, ); } // Only mirror posts that actually got a live route — this is what makes // "draft" (excluded in prod by Docusaurus's own blog plugin) and // ".unpublished/" (outside blog/, never globbed here anyway) fall out // for free, without re-deriving Docusaurus's own draft rules. // `routesPaths` carries the locale's baseUrl (`/fr/blog//`) while `permalinkFor()` // builds a bare `/blog/` — comparing them directly matched nothing under a // non-default locale, so every mirror silently disappeared while the build stayed green. // This is the mirror image of the bug in blog-feed-plugin's getArticleHtml(), where the // prefix was present twice instead of missing once. See TODO 0119. const baseUrl = siteConfig.baseUrl.endsWith("/") ? siteConfig.baseUrl : `${siteConfig.baseUrl}/`; const withBaseUrl = (permalink) => `${baseUrl}${permalink.replace(/^\//, "")}`.replace(/\/$/, ""); const liveRoutes = new Set(routesPaths.map((route) => route.replace(/\/$/, ""))); const live = localeFiltered.filter((post) => liveRoutes.has(withBaseUrl(post.permalink)), ); console.log( `[markdown-export] Mirroring ${live.length}/${posts.length} blog posts to Markdown…`, ); const unknownAll = new Set(); const indexable = []; let failed = 0; for (const post of live) { let markdown; try { const result = await mdxToMarkdown(post.raw, { currentFileDir: post.currentFileDir, projectRoot: siteDir, }); markdown = buildHeader(post, siteConfig) + result.markdown; for (const name of result.unknownComponents) unknownAll.add(name); } catch (err) { // One broken article must never take the whole `yarn build` down — // the "one rule" in degrade.cjs already prevents almost every crash; // this is the last-resort net for whatever it doesn't catch. failed += 1; console.error( `[markdown-export] Failed to degrade ${post.file}: ${err.message}`, ); continue; } const mirrorPath = path.join(outDir, `${post.permalink.replace(/\/$/, "")}.md`); fs.mkdirSync(path.dirname(mirrorPath), { recursive: true }); fs.writeFileSync(mirrorPath, markdown, "utf-8"); if (!post.draft && !post.unlisted) { indexable.push({ ...post, markdown }); } } const seriesFiles = writeSeriesFull(outDir, indexable); writeLlmsTxt(outDir, siteConfig, indexable, seriesFiles); if (unknownAll.size > 0) { console.warn( `[markdown-export] ${unknownAll.size} component(s) had no explicit degrade rule and fell back ` + `to the generic "keep children, drop wrapper" behavior — add a rule in ` + `plugins/markdown-export-plugin/degrade.cjs's COMPONENT_RULES if this is a real gap:\n` + [...unknownAll] .sort() .map((name) => ` - ${name}`) .join("\n"), ); } console.log( `[markdown-export] Done — ${indexable.length} article(s) in /llms.txt, ` + `${live.length - failed}/${live.length} mirrored, ${failed} failed.`, ); }, }; }; ``` Enregistrez-le comme n'importe quel autre plugin : ```javascript title="docusaurus.config.js" plugins: [ "./plugins/markdown-export-plugin/index.cjs", // ... ], ``` > **Auto-hébergé ? Attention au type MIME:** > > `/blog/my-slug` (comme `/blog/docusaurus-llms-txt` pour cet article) est un *répertoire* contenant `index.html`, donc son voisin `/blog/my-slug.md` ne crée aucun conflit de route — celui-là est gratuit. Mais Apache n'a aucun mapping MIME natif pour `.md`, donc sans `AddType text/markdown .md` dans `.htaccess`, le navigateur propose un téléchargement au lieu d'afficher le fichier. ## La seule règle qui rend tout ça sûr {#the-one-rule-that-makes-this-safe} **Un composant inconnu ne fait jamais échouer l'export. Son wrapper est supprimé et ses enfants prennent sa place.** Aucune exception, aucune erreur, jamais. C'est cette unique règle qui permet à une table de dégradation de 900 lignes de vivre dans un pipeline de build sans devenir un risque. Ajoutez un composant dans six mois, oubliez d'écrire sa règle, et le pire cas est que son wrapper disparaisse alors que son contenu survit. Mais un repli silencieux est une fuite lente : le plugin enregistre donc chaque composant qu'il n'a pas reconnu et avertit une fois par build. **C'est cet avertissement, pas la table, qui maintient la couverture à 100 % dans le temps** — la table est figée le jour où vous l'écrivez ; l'avertissement remarque le jour où vous la cassez. ## Rendre le tout découvrable {#making-it-discoverable} Générer un fichier n'est pas le publier : sans lien qui pointe vers lui, personne — crawler ou humain — ne trouve `/llms.txt` ou un bundle de série tout seul. Quatre points d'accroche comblent ce manque, chacun visant un consommateur différent : **1. `llms.txt` référence ses propres bundles.** L'index s'ouvre sur une section « Series (full-text bundles) ». C'est le seul endroit où ils sont référencés. **2. Un `` à l'échelle du site,** dans `headTags`, pour qu'il soit rendu côté serveur sur chaque page — exactement la façon dont on a toujours découvert le RSS : ```javascript title="docusaurus.config.js" { tagName: "link", attributes: { rel: "alternate", type: "text/markdown", href: "https://www.avonture.be/llms.txt", title: "llms.txt — full site index in Markdown, for LLMs and readers", }, }, ``` **3. Un équivalent par article,** pointant vers le miroir de cet article, injecté par un composant minuscule branché sur la page d'article : **src/components/MarkdownAlternate/index.tsx** ```typescript // Points crawlers/tools that scan at this exact // article's plain-Markdown mirror (written by plugins/markdown-export-plugin, // see src/components/CopyAsMarkdown for the human-facing equivalent). Same // discovery pattern as RSS's , // scoped to a single page instead of the whole site — the site-wide // counterpart (pointing at /llms.txt) lives in docusaurus.config.js's // headTags, since that one applies to every page, not just blog posts. import type { JSX } from "react"; import Head from "@docusaurus/Head"; import { useBlogPost } from "@docusaurus/plugin-content-blog/client"; import { translate } from "@docusaurus/Translate"; export default function MarkdownAlternate(): JSX.Element { const { metadata } = useBlogPost(); return ( ); } ``` **4. Un fil d'Ariane dans `robots.txt`.** Il n'existe aucune directive standard pour ça — c'est un commentaire, lu par tout ce qui parse déjà ce fichier et en cherche un. Ça a coûté deux lignes. > **Une limitation, honnêtement:** > > Le lien destiné aux humains « View this series as plain Markdown » sur `/series/` est invisible pour un crawler qui n'exécute pas JavaScript : cette page est une route React Router côté client, donc son lien n'apparaît jamais dans le HTML rendu côté serveur. Les deux balises `` ci-dessus n'ont pas ce problème — et c'est précisément pour ça que les points de découverte qui comptent vivent dans `headTags` et dans un composant, pas dans le corps d'une page. Et puis il y a la partie qu'aucun code ne fera pour vous : **soumettre aux annuaires**. Les bons sont ceux vers lesquels la page officielle de la spec renvoie. J'ai réussi à m'enregistrer sur `llmstxt.site` et `directory.llmstxt.cloud` ; `llmstxthub.com` avait un formulaire de soumission cassé le jour où j'ai essayé. Dix minutes bien investies. ## Conclusion {#conclusion} Ce que j'aime dans le résultat a moins à voir avec l'adoption par les IA que je ne l'imaginais au départ : * **N'importe qui peut faire un `curl` sur un article et l'obtenir en entier.** Pas seulement un bot — un script, un collègue, moi sur une connexion lente — obtient le texte complet, sans HTML, sans JavaScript, sans navigation à nettoyer à la main d'abord. * **Le blog est bot-friendly, que `llms.txt` prenne ou non.** Une version en Markdown pur de chaque article existe maintenant comme effet de bord, indépendamment du sort de cette convention. * **Un fichier Markdown tombe directement dans un éditeur de texte ou dans la fenêtre de contexte d'un LLM.** Plus besoin de tout sélectionner et de supprimer la barre de navigation avant qu'un article soit réellement utilisable ailleurs. * **`/llms.txt` fait aussi office d'inventaire du blog** — chaque article, une ligne chacun, groupé par sujet. Une table des matières que je n'avais pas avant, générée gratuitement. * **Les bundles par série transforment 23 requêtes en une.** Pratique pour un assistant qui récupère du contexte, tout aussi pratique pour un lecteur qui veut une série entière dans un seul fichier, à lire ou à archiver. * **Rien de tout ça n'a besoin d'être retenu.** C'est une étape de build, pas une habitude — et c'est exactement pour ça qu'une convention non prouvée méritait une heure d'essai. Si les 107 accordéons [``](/blog/docusaurus-snippets) fermés qui ont déclenché toute cette histoire sont nouveaux pour vous, ce composant mérite sa propre lecture.