diff --git a/AGENTS.md b/AGENTS.md
index ee49a7c..9b3abe2 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -111,6 +111,8 @@ live Atlassian APIs; property-generated ADF trees; the CommonMark spec suite aga
   `Result<T>` — `{ ok: true; value } | { ok: false; error: ConvertError }` — nothing throws.
   `try/catch` only wrapped tightly around a call that genuinely throws, converted to a result on
   the spot.
+- Only the hard break's inline segment holds a raw newline — every other spelling escapes one or
+  refuses it — which is how the whitespace carry finds a line edge.
 - A readable spelling tried ahead of a general one — the image, the pipe table, a pipe cell —
   returns `string | undefined`, never a `Result`: any failure is the fallback signal, and the
   general form owns the refusal. Refusing there refuses a document the general form spells.
diff --git a/spec/flavour.md b/spec/flavour.md
index 906addd..aeb1d05 100644
--- a/spec/flavour.md
+++ b/spec/flavour.md
@@ -63,7 +63,8 @@ literal bracket. Whitespace at either edge of `[content]`, space or tab, is part
 survives inline parsing. Each section below says whether content is required. `:` opens a
 directive only when the name is followed immediately by `[` or `{`, and `{attrs}` must follow
 `]` (or the name) with no gap — anything else (`10:30`, `:smile:`, a stray `{…}` in text) is
-literal text.
+literal text. An inline directive binds as a unit before bracket matching, the way a code span
+does: a `]` or `(` inside its `{attrs}` is the directive's, never the enclosing content's.
 
 **Container block**:
 
@@ -99,7 +100,8 @@ form, and omits empty `{attrs}` except where the `{` itself claims the directive
 (`:hardBreak{}`).
 
 **Escaping**: the emitter backslash-escapes whatever literal text would otherwise parse as
-directive syntax — the leading `:` of a would-be directive, `]` inside content; outside code
+directive syntax — the leading `:` of a would-be directive, `]` inside content, a `{` right after
+a directive's closing `]`, which would otherwise be read as the attributes it has none of; outside code
 spans and code blocks, a backslash before `:` in input yields a literal colon.
 
 **Malformed directives are error results**, named: an unclosed container at end of input, a body
@@ -214,9 +216,10 @@ One header row plus plain inline cells is a pipe table; anything richer is the d
 (AGENTS.md §4). Precisely: a table emits as a pipe table exactly when the `table`, every row
 and every cell carry no attrs and no marks, the first row is all `tableHeader` and the rest all
 `tableCell`, every row has the header's cell count, and every cell holds exactly one attr-less,
-mark-less paragraph — an empty cell holds one empty paragraph — with no `|` in a code span,
-link destination or link title: backslash escapes are inert in everything the inline layer spells
-as syntax, so pipe form cannot spell that pipe and the table takes the directive form. A pipe table
+mark-less paragraph — an empty cell holds one empty paragraph — with no `|` anywhere the inline
+layer spells as syntax — a code span, a link destination or title, an inline directive's
+attributes: backslash escapes are inert there, so pipe form cannot spell that pipe and the table
+takes the directive form. A pipe table
 parses back to exactly that shape.
 
 ```
diff --git a/src/adf-to-markdown.test.ts b/src/adf-to-markdown.test.ts
index 7bed017..31509cf 100644
--- a/src/adf-to-markdown.test.ts
+++ b/src/adf-to-markdown.test.ts
@@ -228,6 +228,9 @@ test('escapes a literal delimiter that would merge with an emitted one', () => {
   assert.equal(emitted({ text: '`', type: 'text' }, marked('x', { type: 'code' })), '\\``x`\n')
   assert.equal(emitted(marked('x', { type: 'code' }), { text: '`', type: 'text' }), '`x`\\`\n')
   assert.equal(emitted({ text: '!', type: 'text' }, marked('x', { attrs: { href: 'https://example.com/' }, type: 'link' })), '\\![x](https://example.com/)\n')
+  assert.equal(emitted(marked('x', { type: 'underline' }), { text: '{}', type: 'text' }), ':underline[x]\\{}\n')
+  assert.equal(emitted({ attrs: { text: '' }, type: 'status' }, { text: '{color=red}', type: 'text' }), ':status[]\\{color=red}\n')
+  assert.equal(emitted(marked('x', { attrs: { href: 'https://example.com/' }, type: 'link' }), { text: '{}', type: 'text' }), '[x](https://example.com/){}\n')
 })
 
 test('escapes a hyphen underline a hard break would expose', () => {
@@ -377,9 +380,9 @@ test('spells a table as a pipe table only where every row and cell is plain', ()
   const marked = (mark: AdfMark): AdfDocument => table(row(cell('tableHeader', { content: [{ marks: [mark], text: 'l', type: 'text' }], type: 'paragraph' })))
   assert.ok(directive(adfToMarkdown(marked({ attrs: { href: 'https://example.com/?x|y' }, type: 'link' }))))
   assert.ok(directive(adfToMarkdown(marked({ attrs: { href: 'https://example.com/', title: 'a|b' }, type: 'link' }))))
-  const piped = (node: AdfNode): boolean => directive(adfToMarkdown(table(row(cell('tableHeader', { content: [node], type: 'paragraph' })))))
-  assert.ok(piped({ marks: [{ type: 'code' }], text: 'a|b', type: 'text' }))
-  assert.ok(piped({ attrs: { style: 'a|b' }, type: 'status' }))
+  const fallsBack = (node: AdfNode): boolean => directive(adfToMarkdown(table(row(cell('tableHeader', { content: [node], type: 'paragraph' })))))
+  assert.ok(fallsBack({ marks: [{ type: 'code' }], text: 'a|b', type: 'text' }))
+  assert.ok(fallsBack({ attrs: { style: 'a|b' }, type: 'status' }))
   assert.equal(markdown(adfToMarkdown(marked({ attrs: { href: 'https://example.com/x' }, type: 'link' }))), '| [l](https://example.com/x) |\n| --- |\n')
 })
 
@@ -425,6 +428,15 @@ test('spells the directive marks around the longest run they cover', () => {
   assert.equal(emitted(marked('x', { type: 'em' }, underline)), '_:underline[x]_\n')
   assert.equal(emitted(marked('x', underline, { type: 'em' })), ':underline[_x_]\n')
   assert.equal(emitted(marked('a', underline), { marks: [underline], type: 'hardBreak' }, marked('b', underline)), ':underline[a:hardBreak{}b]\n')
+  const link: AdfMark = { attrs: { href: 'https://example.com/' }, type: 'link' }
+  assert.equal(
+    emitted(marked('a', underline, link), { marks: [underline, link], type: 'hardBreak' }, marked('b', underline, link)),
+    ':underline[[a:hardBreak{}b](https://example.com/)]\n',
+  )
+  assert.equal(
+    emitted(marked('a', link), { marks: [link], type: 'hardBreak' }, marked('b', link)),
+    '[a\\\nb](https://example.com/)\n',
+  )
 })
 
 test('refuses a mark directive attribute no spelling holds', () => {
diff --git a/src/markdown-escaping.ts b/src/markdown-escaping.ts
index d72c9a3..ff0820a 100644
--- a/src/markdown-escaping.ts
+++ b/src/markdown-escaping.ts
@@ -18,7 +18,7 @@ const delimiters = ['*', '_', '`', '~']
 
 const asciiPunctuation = /[!"#$%&'()*+,\-./:;<=>?@[\\\]^_`{|}~]/
 const htmlConstructs = [/^<[!?]/, /^<\/?[A-Za-z][A-Za-z0-9-]*(?:[\s/>]|$)/, /^<[^\s<>@]+@[^\s<>@]+>/]
-const inlineDirective = /^:[a-z][A-Za-z0-9]*[[{]/
+const inlineDirectiveOpener = /^:[a-z][A-Za-z0-9]*[[{]/
 const linkOpener = /\](?=[([:])/
 const unicodePunctuation = /[\p{P}\p{S}]/u
 
@@ -115,6 +115,7 @@ function delimiterRuns(segments: readonly InlineSegment[], placements: readonly
 function mergesWithSyntax(scan: string, escapings: readonly (InlineEscaping | undefined)[], index: number): boolean {
   const character = scan.charAt(index)
   if (character === '!') return scan.charAt(index + 1) === '[' && isSyntax(escapings[index + 1])
+  if (character === '{') return scan.charAt(index - 1) === ']' && isSyntax(escapings[index - 1])
   if (!delimiters.includes(character)) return false
   return touchesSyntax(scan, escapings, index, -1) || touchesSyntax(scan, escapings, index, 1)
 }
@@ -157,7 +158,7 @@ function claimsCharacter(scan: string, index: number, inBrackets: boolean, conta
   if (character === '\\') return asciiPunctuation.test(scan.charAt(index + 1))
   if (character === '&') return startsEntityReference(rest)
   if (character === '<') return opensBracketedAutolink(rest) || htmlConstructs.some((construct) => construct.test(rest))
-  if (character === ':') return inlineDirective.test(rest)
+  if (character === ':') return inlineDirectiveOpener.test(rest)
   if (character === '[') return linkOpener.test(rest)
   if (character === '`') return opensCodeSpan(scan, index, escaped)
   if (character === '*' || character === '_' || character === '~') return opensEmphasis(scan, index, escaped)
diff --git a/src/markdown-inline.ts b/src/markdown-inline.ts
index 92e5cce..57d71d9 100644
--- a/src/markdown-inline.ts
+++ b/src/markdown-inline.ts
@@ -9,13 +9,11 @@ import { longestBacktickRun } from './backtick-runs.ts'
 import { serializeCanonicalJson } from './canonical-json.ts'
 import { spellAttributes, spellStringAttribute } from './directive-attributes.ts'
 
-type Brackets = 'directive' | 'link' | 'none'
-
 type InlineContext = {
   atBlockEnd: boolean
-  brackets: Brackets
-  container: LineContainer
+  bracketed: boolean
   path: ConvertErrorPath
+  spansLines: boolean
 }
 
 type InlineRun = { index: number; kind: 'marked'; mark: AdfMark; nodes: AdfNode[] } | { index: number; kind: 'plain'; node: AdfNode }
@@ -47,7 +45,7 @@ export function tryImageLine(alt: string | undefined, href: string, path: Conver
 }
 
 function lineSegments(nodes: readonly AdfNode[], container: LineContainer, path: ConvertErrorPath): Result<InlineSegment[]> {
-  const segments = emitRun(nodes, 0, 0, { atBlockEnd: true, brackets: 'none', container, path })
+  const segments = emitRun(nodes, 0, 0, { atBlockEnd: true, bracketed: false, path, spansLines: container === 'paragraph' })
   if (!segments.ok) return segments
   return success(carryStrippedWhitespace(segments.value))
 }
@@ -69,7 +67,7 @@ function finishLine(segments: readonly InlineSegment[], container: LineContainer
   return success(line)
 }
 
-// spec/flavour.md, Inline nodes: whitespace CommonMark strips.
+// spec/flavour.md, Inline nodes.
 function carryStrippedWhitespace(segments: readonly InlineSegment[]): InlineSegment[] {
   const carried: InlineSegment[] = []
   for (const [index, segment] of segments.entries()) {
@@ -109,7 +107,7 @@ function syntax(text: string): InlineSegment {
 
 function refuseContentAndText(node: AdfNode, path: ConvertErrorPath): Result<null> {
   if ((node.content ?? []).length > 0 || node.text !== undefined) {
-    return failure('unsupported-node-shape', `a ${node.type} node carries neither content nor text`, path)
+    return failure('unsupported-node-shape', `a ${node.type} node holds neither content nor text`, path)
   }
   return success(null)
 }
@@ -165,7 +163,7 @@ function emitHardBreak(node: AdfNode, context: InlineContext, path: ConvertError
   if (unspelled !== undefined) return failure('unspelled-node-attribute', `the hardBreak attribute ${unspelled} has no canonical markdown spelling`, path)
   const empty = refuseContentAndText(node, path)
   if (!empty.ok) return empty
-  if (context.container === 'paragraph' && !context.atBlockEnd && context.brackets !== 'directive') return success([syntax('\\\n')])
+  if (context.spansLines && !context.atBlockEnd) return success([syntax('\\\n')])
   return success([syntax(spellLeafDirective('hardBreak', ''))])
 }
 
@@ -186,11 +184,11 @@ function emitInlineDirective(node: AdfNode, directive: InlineDirective, path: Co
 function emitText(node: AdfNode, context: InlineContext, path: ConvertErrorPath): Result<InlineSegment[]> {
   const unspelled = Object.keys(node.attrs ?? {})[0]
   if (unspelled !== undefined) return failure('unspelled-node-attribute', `the text attribute ${unspelled} has no canonical markdown spelling`, path)
-  if (typeof node.text !== 'string' || node.text === '') return failure('unsupported-node-shape', 'a text node carries no text', path)
-  if ((node.content ?? []).length > 0) return failure('unsupported-node-shape', 'a text node carries no content', path)
+  if (typeof node.text !== 'string' || node.text === '') return failure('unsupported-node-shape', 'a text node holds text', path)
+  if ((node.content ?? []).length > 0) return failure('unsupported-node-shape', 'a text node holds no content', path)
   if (/\r/.test(node.text)) return failure('unspellable-whitespace', 'a text node holds a carriage return CommonMark rewrites', path)
   if (holdsNullCharacter(node.text)) return failure('unspellable-character', 'a text node holds a null character CommonMark replaces', path)
-  const escaping: InlineEscaping = context.brackets === 'none' ? 'backslash' : 'bracketed'
+  const escaping: InlineEscaping = context.bracketed ? 'bracketed' : 'backslash'
   const parts = node.text.split(/(\n+)/).filter((part) => part !== '')
   return success(parts.map((part) => (part.startsWith('\n') ? carriedText(part) : { escaping, text: part })))
 }
@@ -205,7 +203,7 @@ function emitMarkedRun(nodes: readonly AdfNode[], mark: AdfMark, depth: number,
   if (vocabulary === undefined) return failure('unspellable-mark', `no markdown spelling holds the ${mark.type} mark`, path)
   const attributes = spellMarkAttributes(mark, vocabulary, path)
   if (!attributes.ok) return attributes
-  const inner = emitRun(nodes, depth + 1, index, { ...context, brackets: 'directive' })
+  const inner = emitRun(nodes, depth + 1, index, { ...context, bracketed: true, spansLines: false })
   if (!inner.ok) return inner
   return success([syntax(`:${mark.type}[`), ...inner.value, syntax(`]${attributes.value}`)])
 }
@@ -270,7 +268,7 @@ function emitLink(nodes: readonly AdfNode[], mark: AdfMark, depth: number, index
   if (!destination.ok) return destination
   const spelledTitle = title === undefined ? success('') : spellTitle(title, path)
   if (!spelledTitle.ok) return spelledTitle
-  const inner = emitRun(nodes, depth + 1, index, { ...context, brackets: 'link' })
+  const inner = emitRun(nodes, depth + 1, index, { ...context, bracketed: true })
   if (!inner.ok) return inner
   return success([syntax('['), ...inner.value, syntax(`](${destination.value}${spelledTitle.value})`)])
 }