Inside the machine

System prompts extracted from 30+ AI tools. See how they really work — token counts, costs, and techniques.

30 tools 49 prompts $0 API cost
A

Augment

GPT-5
gpt-5

Augment Code — GPT-5 agent prompt.

~3,948
tokens · GPT-4o
2 versions
8/9 techniques
codingagentide
vs
B

Bolt.new

latest

StackBlitz's in-browser Node.js IDE powered by WebContainer.

~5,775
tokens · GPT-4o
2 versions
7/9 techniques
codingweb
vs
C

ChatGPT

GPT-5
2025-11-09

ChatGPT 5 snapshot from November 2025.

~908
tokens · GPT-4o
3 versions
5/9 techniques
chatgeneral
vs
C

Claude

Sonnet 4.6
4.6

claude.ai default system prompt for Sonnet 4.6.

~26,031
tokens · Claude Sonnet 4.6
2 versions
8/9 techniques
chatgeneral
vs
C

Claude Code

2.0

Anthropic's official CLI for Claude. Terse, tool-driven software engineering agent.

~15,084
tokens · Claude Sonnet 4
2 versions
9/9 techniques
codingagentcli
vs
C

Cursor

Agent
2.0

AI code editor built on VS Code. Agent mode — autonomous coding with tool use.

~10,222
tokens · Claude 3.5 Sonnet
5 versions
8/9 techniques
codingagentide
vs
D

Devin

latest

Cognition's autonomous software engineer. Plan/standard-mode workflow.

~9,135
tokens · GPT-4o
1 version
7/9 techniques
codingagent
vs
D

Dia

latest

The Browser Company's Dia assistant.

~4,518
tokens · GPT-4o
1 version
7/9 techniques
chatbrowser
vs
G

Gemini

3.1 Pro
3.1-pro

Google Gemini 3.1 Pro system prompt.

~5,888
tokens · Gemini 1.5 Pro
3 versions
7/9 techniques
chatgeneral
vs
G

GitHub Copilot

Chat
2024-09-30

Microsoft/GitHub's AI pair programmer inside VS Code.

~1,317
tokens · GPT 4
1 version
7/9 techniques
codingchatide
vs
G

Grok

4.2
4.2

xAI's Grok 4.2.

~4,554
tokens · Grok 4.20
2 versions
6/9 techniques
chatgeneral
vs
J

Junie

latest

JetBrains' Junie coding agent.

~1,628
tokens · GPT-4o
1 version
6/9 techniques
codingagentide
vs
K

Kiro

Spec
latest

AWS's developer-focused AI IDE — Spec mode.

~8,407
tokens · GPT-4o
2 versions
8/9 techniques
codingagentide
vs
L

Lovable

latest

AI app builder. Real-time React editing with live preview.

~5,342
tokens · GPT-4o
1 version
5/9 techniques
codingdesignweb
vs
M

Manus

latest

General-purpose agent with planner, knowledge, and data-source modules.

~2,687
tokens · GPT-4o
2 versions
5/9 techniques
agentresearch
vs
M

Microsoft Copilot

2026-03-28

Microsoft Copilot system prompt.

~13,432
tokens · GPT-4o
1 version
9/9 techniques
chatgeneral
vs
N

Notion AI

latest

Notion AI's latest extracted prompt.

~7,211
tokens · GPT-4o
1 version
8/9 techniques
chatproductivity
vs
O

Orchids

latest

Orchids.app coding agent.

~15,462
tokens · GPT-4o
1 version
8/9 techniques
codingagentweb
vs
P

Perplexity

Comet
latest

Perplexity Comet browser assistant.

~6,677
tokens · GPT-4o
2 versions
8/9 techniques
searchchat
vs
Q

Qoder

latest

Qoder coding agent.

~4,900
tokens · GPT-4o
1 version
8/9 techniques
codingagent
vs
R

Replit Agent

Agent
latest

Replit's autonomous programmer. Search-first, tool-driven, workflow-based.

~2,135
tokens · GPT-4o
1 version
6/9 techniques
codingagent
vs
R

RooCode

latest

Open source RooCode agent.

~11,569
tokens · GPT-4o
1 version
9/9 techniques
codingagentide
vs
S

Same.dev

latest

Cloud IDE focused on design-to-code. Uses Bun over NPM, prefers Next.js.

~9,182
tokens · GPT-4o
1 version
9/9 techniques
codingdesignweb
vs
T

Trae

Builder
latest

ByteDance's Trae IDE — Builder mode.

~4,887
tokens · GPT-4o
1 version
9/9 techniques
codingagentide
vs
v

v0

latest

Vercel's generative UI tool. Outputs React/Next.js with shadcn/ui.

~9,545
tokens · GPT-4o
2 versions
8/9 techniques
codingdesignweb
vs
V

VSCode Agent

GPT-5
gpt-5

Microsoft's VSCode Agent — GPT-5 variant.

~6,720
tokens · GPT-4o
2 versions
8/9 techniques
codingagentide
vs
W

Warp

latest

Warp.dev AI terminal.

~3,841
tokens · GPT-4o
1 version
8/9 techniques
agentcli
vs
W

Windsurf

Cascade
Wave 11

Codeium's agentic IDE. Cascade agent with memory system and batched edits.

~3,079
tokens · GPT-4o
2 versions
7/9 techniques
codingagentide
vs
X

Xcode

latest

Apple Xcode AI coding assistant system prompt.

~1,055
tokens · GPT-4o
1 version
4/9 techniques
codingide
vs
Z

Z.ai

latest

Z.ai Code agent.

~2,525
tokens · GPT-4o
1 version
5/9 techniques
codingagent
vs
Learn

Understand what you're looking at

FAQ

Frequently asked questions

What is a system prompt?

A system prompt is the hidden set of instructions given to a large language model before any user message, telling it who it is, what it can do, and how it should behave. It's what turns a generic model like Claude or GPT-4 into a specialized product like Cursor or Perplexity.

How do these system prompts get leaked?

Three common paths: (1) prompt injection attacks where users trick the model into printing its own instructions, (2) inspection of client-side JS bundles for desktop/browser AI tools, and (3) network inspection of the first API request from the tool. Community repositories collect and archive these snapshots.

Are the token counts exact?

For OpenAI-based tools (ChatGPT, GitHub Copilot, VSCode Agent's GPT variants) we use the official tiktoken library — byte-perfect. For Claude, Gemini, Grok, and DeepSeek we use a close approximation of their tokenizers. The order of magnitude is always correct; exact counts vary by a few percent.

Which AI tools are in the directory?

30 tools currently: coding agents (Cursor, Claude Code, Windsurf, v0, Replit, Lovable, Bolt, Devin, Same, Kiro, Trae, Junie, Augment, Orchids, Qoder, RooCode, Z.ai, VSCode Agent, GitHub Copilot, Xcode), chat assistants (ChatGPT, Claude, Gemini, Grok, Microsoft Copilot, Notion AI, Dia), and search/research (Perplexity, Manus). New prompts are added as they surface.

Is it legal to share these prompts?

Legal consensus is that AI system prompts aren't copyrightable in any meaningful sense — they're instructions, not creative work. Most AI companies don't pursue takedowns. WeighMyPrompt hosts community-extracted prompts with clear attribution to source repositories and welcomes requests from tool creators to correct or remove content.

Why is this useful for developers?

Reading real production system prompts is the fastest way to learn prompt engineering. You see how companies structure instructions, handle tool calling, enforce safety, and manage context. Every prompt in the directory includes detected techniques, changelog between versions, and side-by-side comparison against similar tools.

How often is the directory updated?

Prompts are synced from community repositories (x1xhlol, asgeirtj, jujumilk3) on each site update. When a new version of Cursor, Claude, or any tracked tool leaks, we add it to the manifest and re-sync.

Can I submit a prompt?

Yes. Use the Submit form at /system-prompts/submit. Submissions are reviewed before publishing — we don't publish prompts that appear fabricated. Include a source URL (tweet, GitHub, blog post) so we can verify provenance.

Community extracted

System prompts on this page are extracted and shared by the community from public sources. They may be incomplete, outdated, or unverified. WeighMyPrompt does not claim ownership. If you are the creator of a listed tool and want your prompt removed or updated, contact hello@weighmyprompt.com.