Skip to content
ChatGPT & LLMs

Claude System Prompts How-To: Steal Anthropic’s Patterns

Claude system prompts are blowing up again as Anthropic keeps publishing them. Here's how to turn those rules into Project instructions and API systems you can use today.

6 min readBeginner

You’ve typed the same tone rules into Claude for the fifth chat this week and it still drifts. That’s the pain Claude system prompts fix: behavior that lives above the message, not inside it. The topic is hot again because Anthropic keeps publishing the real web/app prompts, and people are stripping them for reusable patterns.

Key takeaway: on claude.ai you never edit Anthropic’s hidden prompt. You stack profile instructions + Project instructions (+ Styles). On the API you own the full system field. Copy structure – short role, sectioned rules, “do this” over long “don’t” lists – not their product-sized dump.

What’s actually going on (quick)

claude.ai and the mobile apps inject a system layer every chat: current date, Markdown-for-code habits, behavior defaults. Those published updates stop at the consumer apps – they do not ride along on API calls (system prompt release notes). Same model family. Different briefing.

That’s why Claude “feels different” inside a third-party app. Someone else wrote the system layer. Mining the published text for filler bans, list rules, and safety phrasing is useful as a mirror. Pasting the whole thing into a Project is not.

Method A vs Method B for Claude system prompts

Match the path to where you already work.

Method A – claude.ai personalization Method B – API system
Control Profile “Instructions for Claude” + per-Project instructions + Styles Full blank-slate system string (or blocks) on each Messages call
Scope Profile = all chats; Project = only that workspace Whatever you send on that request
Default Anthropic web prompt Still underneath; you layer preferences on top Not injected – you define behavior
Best for Writing, research, recurring client work without shipping code Products, agents, tool-heavy workflows, deterministic apps
Hard limit Free accounts: max five Projects (Help Center personalization article, as of July 2026) System tokens count as input on every turn

Start with A if you live in the browser. Ten minutes gets most of the win. B is for when the product is the prompt.

Winner walkthrough: Project instructions that copy Anthropic’s patterns

Account-wide defaults live under your initials → Settings → Instructions for Claude (timezone quirks, “be concise,” jargon you always want). The real system-prompt stand-in is a Project: every chat in that box shares one role, format, and constraint set (Claude Help Center – personalization).

  1. Create a Project named for one job only (e.g. “Client email – Acme”).
  2. Write instructions in short sections, not a novel. Anthropic’s effective context engineering note pushes the “right altitude”: specific enough to steer, not a brittle if/else tree.
  3. Long reference material → Project knowledge. Instructions stay behavior-only.
  4. Styles handle delivery tone (more/less formal) so the instruction block stays about judgment and workflow.
  5. Run three real tasks. Where it ignores you, add one rule per failure – not ten guesses up front.

A tight template that mirrors patterns people keep pulling from the published prompts (answer first, prose over bullet spam, explicit format) without copying Anthropic’s product text:

You are a senior account writer for B2B SaaS.

<priority>
- Lead with the usable answer in the first short paragraph.
- Match the user's energy: short question → short reply; complex brief → structured depth.
- Use plain paragraphs for explanations. Lists only for true inventories or when I ask.
</priority>

<output>
- Subject line options: 3, under 45 characters.
- Body: plain text, max ~120 words unless I request long form.
- End with one suggested CTA sentence.
</output>

<constraints>
- No filler openers.
- Don't invent metrics, customer names, or legal claims.
- If a fact is missing, ask one precise question - then stop.
</constraints>

Building on the API instead? Official prompting best practices still recommend the same moves: role in system, XML-ish sections when prompts get mixed, say what to do instead of only what to avoid.

import anthropic

client = anthropic.Anthropic()
msg = client.messages.create(
 model="MODEL_ID_FROM_YOUR_ACCOUNT", # pick a current ID your workspace lists
 max_tokens=1024,
 system="You are a helpful coding assistant specializing in Python.",
 messages=[{"role": "user", "content": "How do I sort a list of dicts by key?"}],
)

Pro tip: After you draft instructions, paste them into a fresh chat and ask Claude: “Which rules are vague or contradictory? Rewrite at the minimal set that still covers my failures.” Then replace your draft with that rewrite. You’re using the model as a lint pass on its own constitution.

Edge cases that burn people

  • API amnesia: Raw Messages calls do not inherit claude.ai’s published personality. No system string → base manners only.
  • Token tax: System text is input every turn. Community measurements of Claude Code-class stacks climb into the tens of thousands of tokens once tools and feature flags pile on – overhead before your real task starts. Keep your block tight.
  • Cache-friendly steering: On supported newer models, mid-conversation system-role messages can add rules later in messages without rewriting the top-level system and busting a cached prefix (see Anthropic’s mid-conversation system messages docs; confirm model support – this may have changed since early write-ups). Editing the top-level system mid-run is the expensive trap.
  • Free Project ceiling: Five Projects on free means one catch-all box collapses under conflicting rules. Split by workflow, not by mood.

Ever notice how a “perfect” mega-prompt looks sharp on message one and turns mushy by message twelve? That’s attention budget leaking, not you “prompting wrong.”

FAQ

Can I replace Anthropic’s default Claude system prompt on claude.ai?

No. Layer profile instructions, Project instructions, and Styles. Full replacement is an API job.

Should I paste the published Claude system prompt into my Project?

Don’t. It’s built for Anthropic’s full surface – tools, safety routing, app UX – and it’s long. You’ll burn context fighting rules that don’t match your job. Lift patterns (decisive first answer, section tags, concrete output contracts) into a short brief for one workflow. Field guides like Simon Willison’s Claude 4 prompt breakdown beat copy-paste.

Profile instructions or Project instructions – which wins when they conflict?

People assume “more instructions = more control.” Opposite problem: two sources saying different things about tone or format make Claude average them into mush. Keep profile for globals you always want. Put role, deliverable shape, and client-specific bans in the Project. If two rules collide, delete the weaker line from profile – one clear source beats a stack of overrides.

Open claude.ai, create one Project for a task you repeat this week, paste the template above, and run your next real brief inside that Project – not in a blank chat.

Related Tutorials