diff --git a/ui/storybook/prototypes/jev-task-creation/JevTaskComposer.tsx b/ui/storybook/prototypes/jev-task-creation/JevTaskComposer.tsx
index 59f4b3e9269..bbca7cdfe22 100644
--- a/ui/storybook/prototypes/jev-task-creation/JevTaskComposer.tsx
+++ b/ui/storybook/prototypes/jev-task-creation/JevTaskComposer.tsx
@@ -310,7 +310,7 @@ export function JevTaskComposer({
 
       <div className="flex min-h-0 flex-1 flex-col gap-3 overflow-y-auto px-4 pt-4 pb-3">
         {created ? (
-          <p className="whitespace-pre-wrap text-sm text-muted-foreground">{createdPrompt}</p>
+          <p className="whitespace-pre-wrap break-words text-sm text-muted-foreground">{createdPrompt}</p>
         ) : (
           <PromptField value={prompt} expanded={expanded} onChange={setPrompt} onSubmit={create} />
         )}
@@ -512,7 +512,7 @@ function PromptField({
         }
       }}
       placeholder="What should get done?"
-      className="w-full resize-none overflow-hidden bg-transparent text-base leading-relaxed text-foreground outline-none placeholder:text-muted-foreground/50"
+      className="w-full resize-none overflow-hidden break-words bg-transparent text-base leading-relaxed text-foreground outline-none placeholder:text-muted-foreground/50"
     />
   );
 }
diff --git a/ui/storybook/prototypes/jev-task-creation/README.md b/ui/storybook/prototypes/jev-task-creation/README.md
index 30aeb3640a3..6f6d1b96931 100644
--- a/ui/storybook/prototypes/jev-task-creation/README.md
+++ b/ui/storybook/prototypes/jev-task-creation/README.md
@@ -107,6 +107,28 @@ Every task shows an **Estimated cost** table:
 Token counts are estimated at about four characters per token. In production,
 read `usage` from each response instead.
 
+## Titles
+
+The title model gets a short system instruction: name the deliverable, not the
+material to read first; ignore links; name the overall goal when a request has
+phases. Every title, real or simulated, goes through `isUsableTitle` before it is
+shown. A title with a link, more than 10 words, or no text falls back to
+"Untitled task".
+
+The prototype's stand-in (`draftTitle`) approximates this without a model:
+
+- It strips links.
+- It splits the prompt into clauses.
+- It prefers the clause that names a deliverable (fix, build, plan) over reading
+  steps (read, review).
+- It turns questions into "Decide how…".
+- It replaces "it" with the subject named earlier.
+- It cuts at phrase boundaries within nine words.
+
+Story 12c is a real, messy prompt kept as a regression case: a pasted link first,
+reading steps before the deliverable, and two phases. It used to produce a title
+made of the URL.
+
 ## Two flows to compare
 
 - **Suggest-first** (stories 01–11c): review the live suggestions, then create.
diff --git a/ui/storybook/prototypes/jev-task-creation/jev-classifier.ts b/ui/storybook/prototypes/jev-task-creation/jev-classifier.ts
index a94058dcfd1..58d5f06970c 100644
--- a/ui/storybook/prototypes/jev-task-creation/jev-classifier.ts
+++ b/ui/storybook/prototypes/jev-task-creation/jev-classifier.ts
@@ -428,29 +428,129 @@ const CURATED_TITLES: Record<string, string> = {
   "plan the migration": "Plan phased migration of runtime sessions to the new adapter",
 };
 
-const FILLER = /^(hey|hi|please|pls|so|ok|okay|can you|could you|would you|i need you to|i need to|i want to|we need to|we should|let's|lets|i think|it seems like|it looks like)\b[\s,]*/i;
+const FILLER = /^(hey|hi|please|pls|so|ok|okay|also|can you|could you|would you|i need you to|i need to|i want you to|i want to|we need to|we should|let's|lets|i think|it seems like|it looks like)\b[\s,]*/i;
+const URL_PATTERN = /<?\bhttps?:\/\/[^\s>]+>?|\bwww\.[^\s]+/gi;
+/** Verbs that name something to deliver, versus verbs that only name inputs to read first. */
+const DELIVERABLE_VERBS = /^(fix|add|build|create|plan|design|write|draft|ship|implement|migrate|remove|update|investigate|evaluate|research|decide|run|rotate|stop|set up|make|support|tighten|survey|prototype|figure out)\b/i;
+const INPUT_VERBS = /^(read|review|look at|look into|check out|see|open|go through|skim|watch)\b/i;
+const DROP_WORDS = /\b(a|an|the|this|that|these|those|like that|as well|also|just|really|new|some)\b/gi;
+const MAX_TITLE_WORDS = 9;
+/** Phrases that pad a title without adding meaning. */
+const PADDING = /\b(end to end|in general|out there|at once|right now|asap|for now)\b/gi;
+
+function stripFiller(text: string) {
+  let previous = "";
+  let current = text.trim();
+  while (previous !== current) {
+    previous = current;
+    current = current.replace(FILLER, "");
+  }
+  return current;
+}
+
+/** Split a prompt into instruction-sized clauses: sentences, lines, and "and then" / "then" joins. */
+function promptClauses(prompt: string): string[] {
+  return prompt
+    .replace(URL_PATTERN, " ")
+    .split(/(?<=[.!?])\s+|\n+|;\s*|,?\s+(?:and then|then|after that)\s+|\s+and\s+(?=(?:read|review|create|build|plan|fix|add|write|design|make)\b)/i)
+    .map((clause) => stripFiller(clause.replace(/\s+/g, " ").replace(/^(and|but|so)\s+/i, "").replace(/^phase \w+ is to\s+/i, "")))
+    .filter((clause) => clause.split(" ").length >= 2);
+}
+
+/** Pick the clause that says what to deliver; reading and reviewing are inputs, not the task. */
+function deliverableClause(clauses: string[]): string | null {
+  const ranked = clauses
+    .map((clause, index) => {
+      let rank = 0;
+      if (DELIVERABLE_VERBS.test(clause)) rank += 3;
+      // Investigating is real work, but a fix or build in the same prompt is the outcome.
+      if (/^(figure out|investigate|research|survey|explore)\b/i.test(clause)) rank -= 1;
+      if (/\bplan\b/i.test(clause)) rank += 1;
+      if (INPUT_VERBS.test(clause)) rank -= 3;
+      return { clause, rank, index };
+    })
+    .sort((a, b) => b.rank - a.rank || a.index - b.index);
+  return ranked[0] && ranked[0].rank > 0 ? ranked[0].clause : clauses[0] ?? null;
+}
+
+/** The thing a prompt is about, from its first clause ("The issue detail page is slow" -> "issue detail page"). */
+function leadingSubject(clauses: string[]): string | null {
+  const match = clauses[0]?.match(/^(?:the|our|a|an)?\s*([\w-]+(?:\s+[\w-]+){0,3}?)\s+(?:is|are|was|were|keeps?|got|feels?|has|have|looks?|seems?)\b/i);
+  return match ? match[1]!.trim() : null;
+}
+
+/** Turn a question into a decision task ("How should Paperclip support X?" -> "Decide how Paperclip should support X"). */
+function questionToTask(clause: string): string {
+  const how = clause.match(/^how (should|do|does|can|could|would) (?:we\s+)?(.+?)\??$/i);
+  if (how) {
+    const [subject, ...rest] = how[2]!.split(" ");
+    return `Decide how ${subject} ${how[1]!.toLowerCase() === "should" ? "should " : ""}${rest.join(" ")}`;
+  }
+  const should = clause.match(/^should (?:we|i|paperclip)\s+(.+?)\??$/i);
+  if (should) return `Decide whether to ${should[1]}`;
+  const what = clause.match(/^(?:what|which) (.+?)\??$/i);
+  if (what) return `Decide ${what[1]}`;
+  return clause;
+}
+
+/** Compress an instruction into a short imperative title. */
+function compressClause(clause: string, subject: string | null = null): string {
+  let text = questionToTask(clause)
+    // Cut at natural boundaries before any word cap, so titles never stop mid-phrase.
+    .split(/[:(]|,\s*not\b|\s+(?:when|while|if|until|unless|whenever)\s+|\s+-\s+/i)[0]!
+    .replace(/\bit\b/i, subject ?? "it")
+    // "create a plan to build X" -> "Plan X"; "build a plan to use X to build Y" -> "Plan Y".
+    .replace(/^(?:create|build|write|make|draft)\s+(?:a\s+)?plan\s+(?:to|for)\s+(?:use\s+.+?\s+to\s+)?(?:create|build|make|write|add)?\s*/i, "Plan ")
+    .replace(/,?\s*but\s+(?:catered|tailored|aimed|geared)\s+(?:towards|toward|to|for)\s+/i, " for ")
+    .replace(/\s+(?:catered|tailored|aimed|geared)\s+(?:towards|toward|to)\s+/i, " for ")
+    .replace(/\s+(?:so that|so|because|since|which|that will)\s+.*$/i, "")
+    .replace(PADDING, " ")
+    .replace(DROP_WORDS, " ")
+    .replace(/[.!?,:]+$/, "")
+    .replace(/\s+/g, " ")
+    .trim();
+  let words = text.split(" ");
+  if (words.length > MAX_TITLE_WORDS) {
+    // Prefer ending before a trailing phrase ("... worktrees | for local coding agents") over a hard cut.
+    const head = words.slice(0, MAX_TITLE_WORDS);
+    const boundary = head.map((word) => /^(for|to|with|in|on|from|by|across|into)$/i.test(word)).lastIndexOf(true);
+    words = boundary >= 4 ? head.slice(0, boundary) : head;
+  }
+  words = words.map((word, index) => (index === words.length - 1 ? word.replace(/[,;:.]+$/, "") : word));
+  // Never end on a dangling connector.
+  while (words.length > 2 && /^(for|to|of|in|on|with|and|or|by|from|at|based)$/i.test(words[words.length - 1]!)) words.pop();
+  text = words.join(" ");
+  return text.charAt(0).toUpperCase() + text.slice(1);
+}
+
+/**
+ * Checks a title before it is shown, for the stand-in and for the real model alike:
+ * no links, not too long, not cut off. A failing title falls back to a safe one.
+ */
+export function isUsableTitle(title: string): boolean {
+  const words = title.trim().split(/\s+/);
+  // A fresh, non-global copy: `test` on a global regex carries `lastIndex` between calls.
+  const hasUrl = new RegExp(URL_PATTERN.source, "i").test(title);
+  return title.trim().length > 0 && !hasUrl && !/[\/]{2}|\.\w{2,}\//.test(title) && words.length <= 10;
+}
 
 export function draftTitle(prompt: string): string {
   const normalized = prompt.trim().toLowerCase();
   for (const [prefix, title] of Object.entries(CURATED_TITLES)) {
     if (normalized.startsWith(prefix)) return title;
   }
-  let text = prompt.trim().split(/(?<=[.!?])\s|\n/)[0] ?? prompt.trim();
-  let previous = "";
-  while (previous !== text) {
-    previous = text;
-    text = text.replace(FILLER, "");
-  }
-  text = text.replace(/[.!?]+$/, "").replace(/\s+/g, " ");
-  if (text.length > 60) {
-    const cut = text.slice(0, 60);
-    text = cut.slice(0, cut.lastIndexOf(" ") > 30 ? cut.lastIndexOf(" ") : 60);
-  }
-  return text.charAt(0).toUpperCase() + text.slice(1);
+  const clauses = promptClauses(prompt);
+  const clause = deliverableClause(clauses);
+  const title = clause ? compressClause(clause, leadingSubject(clauses)) : "";
+  return isUsableTitle(title) ? title : "Untitled task";
 }
 
-const TITLE_SYSTEM =
-  "You name tasks for an operator dashboard. Reply with only the title: 3 to 8 words, imperative, sentence case, no quotes or trailing punctuation.";
+const TITLE_SYSTEM = [
+  "You name tasks for an operator dashboard.",
+  "Reply with only the title: 3 to 8 words, imperative, sentence case, no quotes or trailing punctuation.",
+  "Name what should be delivered, not the material to read first. Ignore links, URLs, and pasted references.",
+  "If the request has phases, name the overall goal.",
+].join(" ");
 
 /** The Messages API request a server would send to Claude Haiku 4.5 when the task is started. */
 export function buildTitleRequest(prompt: string) {
diff --git a/ui/storybook/stories/jev-task-creation.stories.tsx b/ui/storybook/stories/jev-task-creation.stories.tsx
index 516b30186ae..42e02925290 100644
--- a/ui/storybook/stories/jev-task-creation.stories.tsx
+++ b/ui/storybook/stories/jev-task-creation.stories.tsx
@@ -10,6 +10,9 @@ const PROMPTS = {
   qa: "Before Friday's release, run the smoke tests and verify that invite links still work on authenticated private mode.",
   ambiguous: "The invite emails look wrong, can you sort it out",
   planning: "Plan the migration of agent runtime sessions to the new adapter, in phases, so nothing breaks mid-run.",
+  // A real, messy prompt: a pasted link first, reading steps before the deliverable, and two phases.
+  messy:
+    "https://example.slack.com/archives/C0000000000/p0000000000000000\nRead this Slack thread and review the Meta Muse AI URL. Review the page contents and create a plan to create a page like that, but catered towards the Paperclips connector audience. And then build a plan to use Dota's Paperclip instance to build the new skill based on the new runbook, connector runbook that Dota shared the link to the PR to. And then phase two is to build the website as well.",
 };
 
 const meta = {
@@ -67,7 +70,7 @@ export const Confidence: Story = {
   name: "11 · How sure Jev is",
   args: { initialPrompt: PROMPTS.research },
   play: async ({ canvasElement }) => {
-    const toggle = await within(canvasElement).findByRole("button", { name: /How sure Jev is/ }, { timeout: 4000 });
+    const toggle = await within(canvasElement.ownerDocument.body).findByRole("button", { name: /how sure Jev is/i }, { timeout: 4000 });
     await userEvent.click(toggle);
   },
 };
@@ -75,7 +78,7 @@ export const AssigneeProbabilities: Story = {
   name: "11b · Assignee menu shows Jev's probabilities",
   args: { initialPrompt: PROMPTS.ambiguous },
   play: async ({ canvasElement }) => {
-    const chip = await within(canvasElement).findByRole("button", { name: /^Assignee/ }, { timeout: 4000 });
+    const chip = await within(canvasElement.ownerDocument.body).findByRole("button", { name: /^Owner/ }, { timeout: 4000 });
     await userEvent.click(chip);
   },
 };
@@ -83,7 +86,7 @@ export const ModeProbabilities: Story = {
   name: "11c · Mode menu shows Jev's probabilities",
   args: { initialPrompt: PROMPTS.planning },
   play: async ({ canvasElement }) => {
-    const chip = await within(canvasElement).findByRole("button", { name: /^Mode/ }, { timeout: 4000 });
+    const chip = await within(canvasElement.ownerDocument.body).findByRole("button", { name: /^Mode/ }, { timeout: 4000 });
     await userEvent.click(chip);
   },
 };
@@ -95,7 +98,16 @@ export const InstantEnterToStart: Story = {
   name: "12b · Instant · Enter starts the task, then the title is written",
   args: { flow: "instant", initialPrompt: PROMPTS.feature, titleLatencyMs: 1600 },
   play: async ({ canvasElement }) => {
-    const box = await within(canvasElement).findByRole("textbox", { name: "Describe the task" });
+    const box = await within(canvasElement.ownerDocument.body).findByRole("textbox", { name: "Describe the task" });
+    await new Promise((resolve) => setTimeout(resolve, 900));
+    await userEvent.type(box, "{Enter}");
+  },
+};
+export const MessyRealPrompt: Story = {
+  name: "12c · Messy real prompt · Link first, reading steps, two phases",
+  args: { flow: "instant", initialPrompt: PROMPTS.messy },
+  play: async ({ canvasElement }) => {
+    const box = await within(canvasElement.ownerDocument.body).findByRole("textbox", { name: "Describe the task" });
     await new Promise((resolve) => setTimeout(resolve, 900));
     await userEvent.type(box, "{Enter}");
   },
@@ -108,7 +120,7 @@ export const InstantCreated: Story = {
   name: "13 · Instant · Title and owner arrive",
   args: { flow: "instant", initialPrompt: PROMPTS.feature, latencyMs: 700, titleLatencyMs: 2200 },
   play: async ({ canvasElement }) => {
-    await userEvent.click(await within(canvasElement).findByRole("button", { name: /Start task/ }));
+    await userEvent.click(await within(canvasElement.ownerDocument.body).findByRole("button", { name: /Start task/ }));
   },
 };
 export const JevUnavailable: Story = {