Kit Premier Jour Windsurf

What's in this pack

You installed Codeium Windsurf, opened Cascade for the first time, and now you're hunting for the equivalent of a .cursorrules file, a sensible MCP starter set, and someone honest enough to say which parts of the marketing video are real on day one. That's this pack.

Honest disclaimer up front: the public asset catalog is thin on Windsurf-specific skills today. Most agent-skill production is happening around Claude Code and Cursor. The good news: Windsurf is a VS Code fork, and Cascade reads instructions from a .windsurfrules file plus the open AGENTS.md convention — so most editor-agnostic skills, MCP servers, and prompt patterns drop in without modification. Where a pick is editor-agnostic, we say so.

Who this is for: a developer migrating from VS Code+Copilot, Cursor, or Claude Code who wants Cascade producing real diffs on a real repo by the end of the first afternoon — not someone evaluating IDEs for a team rollout.

Install in this order

1. Windsurf — AI IDE with Cascade Agentic Flows (the IDE itself) — Install Codeium's IDE first, sign in with your account, and verify the free tier gives you Claude + GPT-4 access for Cascade. Confirm Cascade can do a multi-step edit on a throwaway file before installing anything else. If this part is broken, none of the following picks matter.
2. AGENTS.md — Open Format for Coding Agent Instructions (editor-agnostic spec) — Read the spec before you write a single rule. AGENTS.md is the cross-tool convention (Cursor, Windsurf, Claude Code, Aider, Codex CLI all read it or interop with it). Knowing the spec means you stop writing tool-specific rules from scratch every time you switch tools.
3. agentrules — Multi-Provider AGENTS.md Generator (editor-agnostic generator) — Generate .windsurfrules + AGENTS.md + .cursorrules + CLAUDE.md from one source. Pays for itself the first time you try Windsurf on a project that already has a .cursorrules. Don't hand-maintain four files.
4. lnai — Unified AI Config for Claude, Cursor, Copilot & More (editor-agnostic config) — The other half of #3. Where agentrules generates rule files, lnai unifies provider configuration (which model, which keys, which defaults) across the AI coding tools you use side-by-side. Stops you from juggling four ~/.config/ directories.
5. MCP Reference Servers — Official Collection (editor-agnostic MCP) — Windsurf supports MCP. Before installing niche third-party servers, scan the official reference (filesystem, fetch, git, memory, sequential-thinking, time, etc.). Most teams find 60% of what they need here without hunting Smithery.
6. Git MCP — Version Control Server for AI Agents (editor-agnostic MCP) — Specifically the Git MCP. Windsurf's Cascade can read your repo but the MCP server gives it structured access to commit history, blame, branch state, and diffs without parsing shell output. Single highest-leverage MCP for code work.
7. Desktop Commander MCP — Local Terminal + Files (editor-agnostic MCP) — Lets Cascade run shell commands and edit files outside the open workspace. Necessary when you want "upgrade my Node version" or "check disk usage," which Cascade alone can't do. Treat as you'd treat a junior dev with sudo — useful, scope tightly.
8. Playwright MCP — Browser Automation Server (editor-agnostic MCP) — Cascade can drive a browser. Verify the change you just made actually renders, fill forms, capture screenshots, scrape live state. Pairs especially well with Cascade's multi-step flows because the browser step usually reveals the bug the code review missed.
9. Awesome AI System Prompts — 32+ Tool Prompts Revealed (reference, includes Cascade's prompt) — Read Windsurf's actual leaked Cascade system prompt before you write yours. Knowing what the agent already believes about itself prevents you from writing rules that contradict its baseline behavior. Treat this as the manual for "why is Cascade doing X."
10. AI Coding Agents Compared — 2026 Landscape Guide (reference) — Final orientation pick. Once you've got Cascade working, you'll wonder "should I be using Claude Code or Cursor for this task instead?" This is the honest landscape map. Cheaper than five Twitter threads.

How they fit together

Windsurf IDE (#1)
   │
   └─ Cascade reads .windsurfrules + AGENTS.md
         │
AGENTS.md spec (#2) ── agentrules (#3) ── lnai (#4)
   │                       │                  │
   └─ one source ──────────┴──────────────────┘
         ├─ .windsurfrules (Windsurf)
         ├─ AGENTS.md (cross-tool)
         ├─ .cursorrules (Cursor, if you keep it)
         └─ CLAUDE.md (Claude Code, if you keep it)
         │
MCP layer:
   ├─ MCP Reference Servers (#5)  ← scan first
   ├─ Git MCP (#6)                ← daily driver
   ├─ Desktop Commander (#7)      ← shell + files outside workspace
   └─ Playwright MCP (#8)         ← verify in a real browser
         │
Reference shelf:
   ├─ Awesome AI System Prompts (#9)  ← what Cascade actually thinks
   └─ AI Coding Agents Compared (#10) ← when to reach for another tool

The agentrules + AGENTS.md + Git MCP trio is the load-bearing wall. Everything else is decoration without those three. If you only have 30 minutes after installing Windsurf itself, do #3 + #2 + #6 and call it a day.

Tradeoffs you'll hit (Windsurf vs Cursor specifically; Cascade vs Composer)

• Windsurf vs Cursor — Cursor has the bigger asset ecosystem (more .cursorrules examples, more YouTube content, more skills explicitly built for it). Windsurf has Cascade, which is closer to "agent that plans + executes multi-file edits on its own" than Cursor's Composer (which is more "chat that proposes a diff for one task at a time"). If you want a coding partner that drives, Windsurf wins; if you want a fast assistant that responds, Cursor still wins. Picks #3 and #4 explicitly let you run both side-by-side without maintaining duplicate config.
• Cascade vs Composer — Composer is interactive-first (you stay in the loop on every step). Cascade is autonomous-first (it plans, executes, recovers from errors, then surfaces a summary). Cascade is more capable on greenfield refactors. Composer is safer on production codebases where you want to approve each change. Don't fight Cascade's autonomy — give it a tight .windsurfrules (via #3) and let it run.
• Supercomplete vs Copilot autocomplete — Supercomplete predicts multi-line edits, including deletes and refactors of nearby code, not just the next token. Higher hit rate on intent, but more jarring when wrong. Disable it for the first hour while you get used to Cascade; turn it on once your rules file is solid.
• MCP via Windsurf vs MCP via Claude Desktop — Windsurf's MCP integration is newer and quieter about errors. If a server fails to start, you might see no UI feedback. Always test new MCP servers in Claude Desktop or the MCP Inspector first; install in Windsurf only after you've confirmed they work standalone.

Common pitfalls

• Hand-writing .windsurfrules and .cursorrules separately — Don't. Use #3 (agentrules) from day one. Even if you think you'll only use Windsurf, you'll evaluate Claude Code or Cursor within 6 months and wish you had the generator already in place.
• Installing 10 MCP servers at once — Same trap as Claude Code: too many MCPs = slower agent + worse tool selection. Stick to #5–#8 for week one. Only add a server when you hit a specific gap.
• Forgetting Cascade has memory — Cascade tracks per-project memories. If you change your .windsurfrules and Cascade still behaves the old way, clear the project memory — it's caching the previous instruction set.
• Trusting the leaked system prompt as gospel — #9 is invaluable, but vendors update system prompts silently. Treat it as "what Cascade believed last month," not "what Cascade believes now." Re-check quarterly.
• Skipping the landscape map (#10) because you already picked Windsurf — You picked Windsurf for this project. Six months from now, a different task will fit Claude Code or a CLI agent better. Knowing the landscape isn't a sign of indecision — it's how you avoid forcing every task into the one IDE you happen to have open.