diff --git a/CLAUDE.md b/CLAUDE.md
index 6a54fd36..76a722e1 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -200,6 +200,9 @@ Big tool output accumulates in the conversation; when context fills, Claude Code
 compacts it and DETAILS get lost (exact node IDs, values, what was tried), which
 shows up as confidently-wrong recall ("hallucinated" IDs). Keep context lean:
 
+- **Pull ONE section of this guide instead of reading all of it.** The whole file is
+  ~10,200 tokens; `figma-cli docs jsx-syntax` is ~930. `figma-cli docs` lists the
+  topics with their token cost. Read the file whole only when you truly need it all.
 - **`verify` saves the PNG to disk by default** and returns just dimensions, instead
   of dumping a base64 image (thousands of tokens) into context. Use `--save <path>`
   only to pick a custom path; avoid `--base64` unless a script needs the inline data.
@@ -752,6 +755,8 @@ figma-cli recreate-url "URL"         # Recreate webpage in Figma
 figma-cli screenshot-url "URL"       # Screenshot webpage
 figma-cli daemon status              # Check daemon
 figma-cli daemon restart             # Restart daemon
+figma-cli docs                       # List this guide's topics + token cost
+figma-cli docs jsx-syntax            # Print ONE section (~930 tok, not ~10,200)
 ```
 
 For eval patterns, layout examples, and Safe Mode templates, see REFERENCE.md.
diff --git a/src/commands/docs.js b/src/commands/docs.js
new file mode 100644
index 00000000..cbeb3b5f
--- /dev/null
+++ b/src/commands/docs.js
@@ -0,0 +1,67 @@
+// Commands: docs — read the usage guide one section at a time.
+import chalk from 'chalk';
+import { existsSync, readFileSync } from 'fs';
+import { fileURLToPath } from 'url';
+import { dirname, join } from 'path';
+import { program } from '../lib/cli-core.js';
+import { parseSections, matchSections } from '../lib/doc-sections.js';
+
+const ROOT = join(dirname(fileURLToPath(import.meta.url)), '..', '..');
+
+// The guide normally sits at the repo root as CLAUDE.md. A checkout that wants
+// a short CLAUDE.md for repo instructions moves the guide to
+// docs/FIGMA-USAGE.md — that one is checked FIRST, because where it exists it
+// is the guide and the root CLAUDE.md is something else.
+const GUIDE_CANDIDATES = ['docs/FIGMA-USAGE.md', 'CLAUDE.md'];
+
+function loadGuide(explicitPath) {
+  const candidates = explicitPath ? [explicitPath] : GUIDE_CANDIDATES;
+  for (const rel of candidates) {
+    const path = rel.startsWith('/') ? rel : join(ROOT, rel);
+    if (existsSync(path)) return { path, rel, markdown: readFileSync(path, 'utf8') };
+  }
+  return null;
+}
+
+program
+  .command('docs [topic]')
+  .description('Print ONE section of the usage guide (no topic = list them). Cheaper than reading the whole guide.')
+  .option('-f, --file <path>', 'read a different markdown file')
+  .action((topic, options) => {
+    const guide = loadGuide(options.file);
+    if (!guide) {
+      console.error(chalk.red(`Usage guide not found (looked for ${(options.file ? [options.file] : GUIDE_CANDIDATES).join(', ')})`));
+      process.exit(1);
+    }
+
+    const sections = parseSections(guide.markdown);
+
+    if (!topic) {
+      const total = sections.reduce((sum, s) => sum + s.tokens, 0);
+      console.log(chalk.bold(`\n  ${guide.rel} — ${sections.length} topics, ~${total} tokens in full\n`));
+      for (const s of sections) {
+        console.log(`  ${chalk.cyan(s.slug.padEnd(28))} ${chalk.gray(String(s.tokens).padStart(5) + ' tok')}  ${s.heading}`);
+      }
+      console.log(chalk.gray(`\n  figma-cli docs <topic>   prints one section instead of all ~${total} tokens\n`));
+      return;
+    }
+
+    const hits = matchSections(sections, topic);
+
+    if (hits.length === 0) {
+      console.error(chalk.yellow(`No topic matches "${topic}".`));
+      console.error(chalk.gray('  Available: ') + sections.map((s) => s.slug).join(', '));
+      process.exit(1);
+    }
+
+    // Several matches: name them instead of guessing which one was meant.
+    if (hits.length > 1) {
+      console.error(chalk.yellow(`"${topic}" matches ${hits.length} topics — pick one:`));
+      for (const s of hits) {
+        console.error(`  ${chalk.cyan(s.slug.padEnd(28))} ${chalk.gray(s.tokens + ' tok')}  ${s.heading}`);
+      }
+      process.exit(1);
+    }
+
+    process.stdout.write(hits[0].body);
+  });
diff --git a/src/lib/command-map.js b/src/lib/command-map.js
index 34e08270..c37e6adf 100644
--- a/src/lib/command-map.js
+++ b/src/lib/command-map.js
@@ -7,7 +7,7 @@ export const ALL = [
   'setup', 'variables', 'daemon', 'tokens', 'gradient', 'create', 'url-tools',
   'config', 'canvas-ops', 'render', 'export-eval', 'analyze', 'a11y',
   'node-ops', 'slots', 'figjam', 'variants', 'misc', 'extract', 'rules',
-  'snapshot', 'spec', 'instantiate', 'init', 'motion',
+  'snapshot', 'spec', 'instantiate', 'init', 'motion', 'docs',
 ];
 
 /**
@@ -112,4 +112,5 @@ export const COMMAND_MODULES = {
   instantiate: ['instantiate'],
   'init-agent': ['init'],
   motion: ['motion'],
+  docs: ['docs'],
 };
diff --git a/src/lib/doc-sections.js b/src/lib/doc-sections.js
new file mode 100644
index 00000000..27999336
--- /dev/null
+++ b/src/lib/doc-sections.js
@@ -0,0 +1,89 @@
+/**
+ * Reading ONE section out of the usage guide instead of the whole file.
+ *
+ * The guide is ~10k tokens. An agent told to "read the docs first" pays that
+ * on every session, even when the task needs one 900-token section. Splitting
+ * it by `##` heading lets a caller pull just what it needs.
+ *
+ * Pure string work, so the matching is unit-tested without touching disk.
+ */
+
+/** Heading text → a stable slug: lowercase words, everything else a dash. */
+export function slugify(heading) {
+  return heading
+    .toLowerCase()
+    .replace(/\([^)]*\)/g, ' ')   // drop parenthetical asides
+    .replace(/[^a-z0-9]+/g, '-')
+    .replace(/^-+|-+$/g, '');
+}
+
+/**
+ * Split a markdown document at its `##` headings.
+ *
+ * `###` and deeper stay inside their parent section — they are subsections of
+ * the topic, not topics of their own. Text before the first `##` (the title) is
+ * dropped: it carries no topic.
+ *
+ * @returns {{heading: string, slug: string, body: string, tokens: number}[]}
+ */
+export function parseSections(markdown) {
+  const lines = markdown.split('\n');
+  const sections = [];
+  let current = null;
+
+  for (const line of lines) {
+    const isTopLevel = /^##\s+/.test(line) && !/^###/.test(line);
+    if (isTopLevel) {
+      if (current) sections.push(current);
+      const heading = line.replace(/^##\s+/, '').trim();
+      current = { heading, slug: slugify(heading), body: line + '\n' };
+      continue;
+    }
+    if (current) current.body += line + '\n';
+  }
+  if (current) sections.push(current);
+
+  // ≈ tokens, for the listing. Bytes/4 is the usual rough rule; it only has to
+  // be good enough to tell a 900-token section from a 2000-token one.
+  for (const s of sections) {
+    s.body = s.body.replace(/\n+$/, '\n');
+    s.tokens = Math.round(s.body.length / 4);
+  }
+  return sections;
+}
+
+/**
+ * Find the sections a query asks for.
+ *
+ * Ranked, because a query should not have to be exact:
+ *   1. slug is exactly the query          "jsx-syntax"  → JSX Syntax
+ *   2. a slug word is exactly the query   "jsx"         → JSX Syntax
+ *   3. slug contains the query            "pitfall"     → Critical Pitfalls
+ *   4. heading contains the query          "render"     → JSX Syntax (render command)
+ *
+ * Returns every section at the best tier that matched, so an ambiguous query
+ * ("token" → Design Tokens + Token Hygiene) can be reported rather than
+ * silently resolved to whichever came first.
+ */
+export function matchSections(sections, query) {
+  const q = slugify(String(query || ''));
+  if (!q) return [];
+
+  // "token" should find "Design Tokens" as readily as "Token Hygiene", so a
+  // trailing plural s is ignored on both sides of a word comparison.
+  const singular = (w) => w.replace(/s$/, '');
+  const qs = singular(q);
+
+  const tiers = [
+    (s) => s.slug === q,
+    (s) => s.slug.split('-').some((w) => singular(w) === qs),
+    (s) => s.slug.includes(q),
+    (s) => slugify(s.heading).includes(q) || s.heading.toLowerCase().includes(query.toLowerCase()),
+  ];
+
+  for (const test of tiers) {
+    const hits = sections.filter(test);
+    if (hits.length) return hits;
+  }
+  return [];
+}
diff --git a/tests/doc-sections.test.js b/tests/doc-sections.test.js
new file mode 100644
index 00000000..78ddcd6a
--- /dev/null
+++ b/tests/doc-sections.test.js
@@ -0,0 +1,94 @@
+// Unit tests for the usage-guide section splitter (pure, no disk access).
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { parseSections, matchSections, slugify } from '../src/lib/doc-sections.js';
+
+const DOC = `# figma-ds-cli
+
+Intro text that belongs to no topic.
+
+## Quick Reference
+
+| a | b |
+
+## JSX Syntax (render command)
+
+Use <Frame>.
+
+### Nested heading stays here
+
+More JSX detail.
+
+## Critical Pitfalls
+
+Don't do that.
+
+## Design Tokens
+
+tokens
+
+## Token Hygiene (keep context lean)
+
+hygiene
+`;
+
+test('splits on ## and keeps ### inside its parent section', () => {
+  const sections = parseSections(DOC);
+  assert.deepEqual(sections.map((s) => s.heading), [
+    'Quick Reference',
+    'JSX Syntax (render command)',
+    'Critical Pitfalls',
+    'Design Tokens',
+    'Token Hygiene (keep context lean)',
+  ]);
+  const jsx = sections.find((s) => s.slug === 'jsx-syntax');
+  assert.ok(jsx.body.includes('### Nested heading stays here'));
+  assert.ok(jsx.body.includes('More JSX detail.'));
+});
+
+test('text before the first ## is dropped', () => {
+  const joined = parseSections(DOC).map((s) => s.body).join('');
+  assert.ok(!joined.includes('Intro text that belongs to no topic.'));
+});
+
+test('slugify drops parenthetical asides', () => {
+  assert.equal(slugify('JSX Syntax (render command)'), 'jsx-syntax');
+  assert.equal(slugify('Motion (Figma Animation, Config 2026 Beta)'), 'motion');
+});
+
+test('a single word matches the section it names', () => {
+  const hits = matchSections(parseSections(DOC), 'jsx');
+  assert.equal(hits.length, 1);
+  assert.equal(hits[0].heading, 'JSX Syntax (render command)');
+});
+
+test('a partial word still matches', () => {
+  const hits = matchSections(parseSections(DOC), 'pitfall');
+  assert.equal(hits.length, 1);
+  assert.equal(hits[0].heading, 'Critical Pitfalls');
+});
+
+test('an ambiguous query returns every candidate rather than guessing', () => {
+  const hits = matchSections(parseSections(DOC), 'token');
+  assert.deepEqual(hits.map((s) => s.heading), ['Design Tokens', 'Token Hygiene (keep context lean)']);
+});
+
+test('an exact slug beats the sections that merely contain it', () => {
+  const hits = matchSections(parseSections(DOC), 'design-tokens');
+  assert.equal(hits.length, 1);
+  assert.equal(hits[0].heading, 'Design Tokens');
+});
+
+test('an unknown query matches nothing', () => {
+  assert.deepEqual(matchSections(parseSections(DOC), 'kubernetes'), []);
+});
+
+test('an empty query matches nothing', () => {
+  assert.deepEqual(matchSections(parseSections(DOC), ''), []);
+});
+
+test('each section carries a token estimate', () => {
+  const quick = parseSections(DOC).find((s) => s.slug === 'quick-reference');
+  assert.ok(quick.tokens > 0);
+  assert.equal(quick.tokens, Math.round(quick.body.length / 4));
+});