Claude Haiku 5.5 is told to write reports without bullets, and to check its memory first
The system prompt Anthropic publishes for Claude Haiku 5.5 opens on an instruction that appears in none of the other 21 published prompts: look in memory before answering, and save what is worth keeping before the turn is over. The document, headed October 7 and read on October 13, makes it the model's first instruction.
Anthropic's own diagram of the three layers of safeguards it puts around its agents. Image: Anthropic
The wording is specific about when to look. "When a question is about the person or their world, Claude checks the memory_listing before answering from the conversation alone, and opens any file whose description suggests it holds something the reply needs. When Claude learns something lasting about the person, it saves it to memory before the turn ends." Fetching all 22 prompts and searching them shows memory_listing in this one and nowhere else.
What the line does not do is describe the feature. It tells you a listing and files exist in the app's context, and that is all. It says nothing about what Anthropic's consumer memory stores, how long it keeps it, or whether a person can switch it off, and the memory stores documented for the API are a separate, developer-facing product.
Reports get no bullet points at all
The formatting instruction is the blunter half. For reports, documents, technical documentation and explanations, the prompt says Claude writes prose without bullets, numbered lists or excessive bolding, and then removes the wriggle room in brackets: "i.e. its prose should never include bullets, numbered lists, or excessive bolded text anywhere", unless the person asks for a list or a ranking. Inside prose, it adds, lists should read naturally as "some things include: x, y, and z", with no newlines. Elsewhere, formatting is allowed only when asked for or when the content is multifaceted enough to need it, and a bullet point has to run to at least one or two sentences.
That wording is not new to Haiku 5.5, and Haiku 5.5 is not the document that changed. The same sentence appears verbatim in 14 of the 22 published prompts, the Claude Opus 5.5 prompt of September 22 among them. The one that drops it is the Claude Sonnet 5.5 prompt of September 28, which says the opposite: "Claude uses lists and bullet points when asked to or when the content is multifaceted enough that they help with clarity. Claude can use bullet points and markdown formatting to make outputs more readable."
So two of the four current models have published prompts that disagree about bullet points, and the newest of them keeps the older house wording while the September outlier is Sonnet 5.5, the model that can also hand your conversation to Sonnet 5 over something you never typed. Anthropic gives no reason for the difference in either document.
The prompt also tells Haiku what it is
Anthropic has the model introduce itself in plain terms: "This iteration of Claude is Claude Haiku 5.5. Claude Haiku 5.5 is the fastest model for quick questions." The line-up it is told to name is Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5, and the cutoff it is given is the end of June 2026, "past which it can't answer reliably", at which point the prompt tells it to say so and point the person to web search.
Two limits apply to the lot. These prompts cover claude.ai and the Claude iOS and Android apps, and Anthropic's overview page says the updates do not apply to the Claude API, so none of it governs what a developer's request returns. And an instruction is not a result: the no-bullets rule is what Anthropic tells the model, not a measurement of what it writes. Haiku 5.5 is the cheap, fast tier of the four, and it arrived on October 7, the date on this prompt.
