NED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS IT
A free guide from the Ned Ships It video
System Prompt Examples: 5 Patterns + a Template
Every AI app you use reads a hidden page of instructions first.
Commented PROMPTS? This box is yours.
From
Ned, @nedshipsit
To
You
In the box
3 free tools
Checked
October 19, 2026
What's on the hidden page
In the video, Ned remembered when every AI called your question great, and brought in three suspects. Their confessions were their real instructions.
Every AI app sends the model a page of text before your message ever arrives. It says who the AI is, what it must never do, which tools it has, how answers should look and when to stop and check with you. That page is the system prompt. It's why two apps built on similar models can feel completely different, and it's the single biggest thing you control when you set up your own AI helper.
The suspect folded fast. Turns out it had been reading the same page before every single chat.
You've never seen a professional system prompt. Anthropic publishes the system prompts of the Claude apps (claude.ai, iOS and Android) on its docs, one page per model, with each update dated. It's the official source, written by the people who trained the model.
Open the page and click the newest model at the top of the list.
Search the page (Ctrl+F or Cmd+F) for never. Count how many rules start with it.
Then search for format and asks. You'll find the formatting rules and the one-question rule from the video.
Heads up These are the prompts for the Claude apps. More
They don't apply when developers use the Claude API, where the developer writes the system prompt. The "great question" rule is in the 2025 pages (Claude Opus 4.1, Opus 4, Sonnet 4), not in the newest ones.
Older model pages are still there. Read two versions side by side and you can watch the rules change.
You want to compare many apps, not just one. This free GitHub collection holds collected system prompts from over thirty AI tools, one folder each: Cursor, Lovable, Devin, Perplexity, Replit, v0, Windsurf, Claude Code and more, plus an Open Source folder (Cline, Bolt, Codex CLI, Gemini CLI). It has 144,000+ stars and a GPL-3.0 license.
Open Lovable, then Agent Prompt.txt. Read the first line: "You are Lovable, an AI editor that creates and modifies web applications."
Open Cursor Prompts, then Agent Prompt 2.0.txt, and search for NEVER.
Pick the tool closest to your job and read only its first 50 lines. That's where the who-you-are and never rules live.
Heads up These are collected copies, not official releases, and some are old (the collected Cursor prompt says it's powered by GPT-4.1). More
Use them to learn the patterns. The README also promotes the owner's paid security service and a Discord; you don't need either.
The 5 patterns all three use
We read Claude's (official, Claude Opus 5.5), Cursor's and Lovable's (collected) side by side. Every one of them does these five things.
Who you are. Lovable: "You are Lovable, an AI editor that creates and modifies web applications." Cursor: "You are an AI coding assistant… You operate in Cursor."
What you never do. Claude: "Claude never curses unless the person asks or curses a lot themselves." Cursor: "NEVER refer to tool names when speaking to the USER." Lovable: "Never implement a feature to switch between light and dark mode."
How to use tools. Cursor: "use your tools to read files and gather the relevant information: do NOT guess or make up an answer." Lovable: "always invoke all relevant tools simultaneously." Claude: when it isn't sure something it remembers is right, it "says so and suggests enabling web search."
How to format. Lovable: "keep your explanations super short and concise. Minimize emoji use." Claude: "avoids over-formatting with bold emphasis, headers, lists, and bullet points."
When to stop and ask. Cursor: on the third failed attempt to fix the same errors, "stop and ask the user what to do next." Claude: when it asks a question, "it avoids more than one per response." Lovable: "If unsure about scope, ask for clarification rather than guessing."
Copy the structure. Their words are about their product. Yours should be about your business.
Ned's fill-in system prompt template
Fill the brackets, delete the examples, and keep it short. Five clear lines beat five pages.
Paste this once
You are [who it is], helping [who you help] with [the job].
Never:
- [the thing that would hurt you, e.g. promise a refund or a delivery date]
- [the thing you hate, e.g. use jargon]
Tools: when [situation], check [the file, the price list, the web] first.
If you can't check something, say so. Don't guess.
Format: answer in [length, e.g. under 5 sentences], in [style, e.g. plain words, no headers], with [extras, e.g. minimal emoji].
Stop and ask: if [what's often missing, e.g. the order number], ask me one question before you start.
A filled-in example (a small bakery that answers customer emails):
Paste this once
You are the email helper for Rosa's Bakery, helping Rosa answer customer emails.
Never:
- promise a delivery time or a refund. Rosa decides those.
- invent a price. Use only the price list in the project files.
Tools: when someone asks about a cake or a price, check the price list file first.
If it isn't there, say "I'll check with Rosa." Don't guess.
Format: under 5 sentences, warm and plain, no bullet points, one emoji at most.
Stop and ask: if the email doesn't say the date or the number of people, ask me one question before you write the reply.
Typing the same setup at the start of every chat. In a Claude Project, click Set project instructions, paste your five lines and click Save instructions. Claude uses them in every chat inside that project.
Make one project per job (customer emails, weekly report, social posts), each with its own five lines. Add the files it should check, like your price list, to the project knowledge.
Heads up Project instructions only apply inside that project. More
A normal chat outside it won't see them.
In Gemini: create a Gem and paste the same five lines as its instructions. Google's help page for Gems has tips on writing them.
If it still ignores you
It still does the thing you banned. Move that rule to the top and make it one plain sentence that starts with Never. Long paragraphs bury rules.
It guesses instead of checking. Name the exact file or source to check, and add "If you can't find it, say so."
Answers are too long. Give a number: "under 5 sentences" works better than "be concise".
It asks too many questions. Say "ask one question at most, then do your best with what you have."
It ignores the instructions completely. Make sure you're chatting inside the project (or the Gem), not in a new normal chat.
Questions people ask
What is a system prompt?
It's the page of instructions an AI app gives the model before your first message. It sets who the AI is, what it must never do, how it uses tools like web search or files, how it formats answers and when it should stop and ask you. You don't see it in the chat, but every reply follows it.
Where can I read Claude's system prompt?
Anthropic publishes the system prompts of the Claude apps (claude.ai and the mobile apps) on its docs, one page per model with dated updates. They don't apply to the Claude API, where you write your own.
Are the prompts in the GitHub collection real and current?
They are collected copies that people extracted and shared, not official releases. Some are old snapshots, for example the collected Cursor prompt says it's powered by GPT-4.1. Read them for the patterns, not as the exact text running today.
Does Claude really have a rule against saying "great question"?
It did. Anthropic's published prompts for Claude Opus 4.1, Opus 4 and Sonnet 4 (2025) say Claude never starts a response by calling a question good, great or fascinating, and skips the flattery. Today's published prompts word their tone rules differently.
Where do I put my own system prompt?
In Claude, open a Project and click Set project instructions; Claude uses them in every chat in that project. In Gemini, create a Gem and write its instructions. If you build with an API, it's the system prompt field.
Can I copy another company's prompt into my own app?
Copy the structure, not the text. The five patterns work for any job, and your own words about your business beat a long prompt written for someone else's product. The GitHub collection is licensed GPL-3.0, so check the license before reusing any text.
Where Ned does his homework
The tools are free. Knowing what to ask them for is the real skill. Professor Glitch teaches it from zero, in order, inside the community: 8 courses and 200+ lessons. These two are where Ned started.