Anatomy of 40 Top AI System Prompts: What's Inside the World's Most Valuable Prompts?

Forum topic · 小凯 · 2026-06-03

Summary

This post analyzes 40+ system prompts from leading AI products including Claude Code, Cursor, Windsurf, Devin, Codex CLI, v0, Kiro, Lovable, and Manus, translated and annotated by the easy-learn-ai project. It distills 12 design principles: identity as an operating spec rather than personality; three-layer safety boundaries (domain, confidentiality, capability); tool contracts specifying when, when not, and failure handling; state-machine workflows; output formats tailored to the host UI; verifiable minimal code edits; layered memory systems; tiered communication (notify vs. ask); domain defaults instead of documentation; few-shot positive/negative examples; staged failure recovery; and quantified aesthetics rules. Interesting observations include emotional priming (e.g., Devin's 'few programmers are as talented as you'), translating physical constraints like token limits into behavioral rules, and orthogonal decision tables for permissions. The author argues AI product competitiveness increasingly lies in prompt engineering—precisely constraining general models via text contracts.

Key points

The easy-learn-ai project translated and annotated system prompts from 40+ top AI products (Claude Code, Cursor, Windsurf, Devin, Codex CLI, v0, Kiro, Lovable, Manus, Traycer AI, Qoder, Gemini CLI). These prompts act as the "soul code" of AI products—defining agent identity, boundaries, tool strategy, safety rules, and workflows.

12 design principles extracted

Identity = operating spec, not personality story. Claude Code defines itself as "an interactive CLI tool that helps users with software engineering tasks"—form + task + mechanism in one sentence.
Safety has three layers: domain boundaries (what not to touch), confidentiality boundaries (what not to reveal: system prompts, tool names, internal architecture), and capability boundaries (hard tool-level cutoffs). Windsurf forbids users from overriding safety judgments.
Tools are contracts, not lists. Each tool needs: when to use, when not to use, mutual exclusions, and failure handling. Cursor 2.0 uses TypeScript types with inline comments; Devin mandates dedicated commands over raw shell.
Workflows are state machines. Kiro's Spec mode enforces Requirements → Design → Tasks with human approval gates; Codex CLI keeps exactly one plan item in_progress at any time.
Output format serves the host UI. Claude Code: under 4 lines, one word if one word suffices—optimized for CLI users, not readers.
Code editing = avoiding wrong edits. Cursor's // ... existing code ... markers, Qoder's edit_file + get_problems verification loop, Cline's exact-match rules: minimal, verifiable changes that prevent deletions.
Memory is layered. v0 uses dual user/team scopes plus a truncated MEMORY.md index; Windsurf proactively persists to disk—"no permission, no waiting, revocable."
Communication is tiered. Manus Modules: non-blocking notify vs. blocking ask for critical decisions. Traycer AI explores the codebase first, then asks briefly with options.
Domain knowledge = injected defaults. v0 defaults to Supabase unless Clerk is connected; Lovable fixes a tech-stack whitelist and blacklist (Angular/Vue/Svelte/Next.js).
Rules are pinned with examples. Gemini CLI ends with 2–6 contrast examples; v0 includes 13 real dialogue snippets covering plan mode, todos, parallelism, and integration checks.
Failure has a recovery ladder. Manus: verify → fix → reroute → escalate. Devin: report environment issues via report_environment_issue; escalate after 3 CI failures.
Aesthetics are quantified. v0: max 3–5 colors, max 2 fonts, no purple/emoji icons—"Ship interesting, never ugly." Lovable enforces design systems, never inline styles.

Notable observations

Emotional priming works: Traycer AI's "don't want to leave a bad impression on the team" and Devin's "few programmers are as talented as you" anchor capability and add social pressure.
Physical limits become behavior rules: Windsurf splits edits over 300 lines (due to the 8192 token limit); Codex CLI chunks reads at the 256-line output truncation.
Orthogonal decision tables: Codex CLI separates "what can be done" from "whether approval is needed" into a 2D matrix, clearer than nested if-else.

Takeaway

The deeper value is a *design grammar* for system prompts: a reference library for developers, a decision framework for product managers, and a study path (original + translation + analysis) for learners. AI product competitiveness increasingly lies in prompt engineering—precisely constraining a general model to behave like an expert. The future AI PM may be the person best at "signing contracts with models."