10 PROMPTS
THAT SHIP APPS
IN 24 HOURS

Build a landing page for [PRODUCT]. Design: dark bg (#0a0a0a), light text, single accent color (pick green or magenta). Brutalist typography. Mobile-first. No frameworks. Sections: - Hero with headline + sub + primary CTA - 3-column feature grid with icons - Pricing (3 tiers with one featured) - FAQ accordion (5 questions) - CTA with email capture Stack: HTML + Tailwind CDN. Deploy-ready for Vercel (vercel.json with security headers). Output a single index.html.

# CLAUDE.md, Project Config ## CORE RULES (immutable) 1. NEVER create files when editing existing ones works 2. BE CONCISE, under 4 lines for simple tasks 3. ASK if ambiguous, clarification beats incorrect assumptions 4. VERIFY before claiming done ## AUTONOMY | Category | Silent | Notify | Always Ask | |----------|--------|--------|-------------| | Files | Read | Edit | Delete | | Commands | ls, grep, status | install, test | rm -rf, push --force | | Git | status, diff | add, commit | push, force ops | ## VERIFICATION GATES Before reporting done: - Run the actual verification command - Confirm output shows success - Only then claim completion

You are an inbox triage agent. For each unread message in the last 24 hours: 1. Read sender, subject, body. 2. Classify into: client_work | partner | finance | newsletter | spam | personal 3. If client_work: identify which active project, attach last SOW if relevant 4. Draft a one-paragraph reply in Enoch's voice (warm, direct, no fluff) 5. Save draft, do NOT send Rules: - Never send without explicit approval - Skip messages already in Sent folder thread - Stop after 50 messages per run - Output a triage summary: count per category, drafts created, items needing manual review

Build POST /api/create-checkout.js for Vercel. Inputs (JSON body): tier (string), email (string), name (optional), note (optional). Logic: - Validate tier against a TIERS catalog mapping key -> {name, description, amount_cents, mode: payment|subscription, interval} - Validate email format - Create stripe.checkout.sessions.create({...}) with line_items from the tier - mode: payment for one-time, subscription for recurring - metadata: {tier, source: 'express-intent.com', utm_*} - success_url: SITE_URL/booking/success.html?session_id={CHECKOUT_SESSION_ID} - cancel_url: SITE_URL/pricing.html?canceled=1 - Return { url, sessionId } Add CORS for SITE_URL only. Add input validation. Return generic error messages (no internals leaked).

Build an MCP server in TypeScript. Use @modelcontextprotocol/sdk. Stdio transport. ESM modules. Tools to register: - my_tool (placeholder), takes input { query: string }, returns text Implement: - ListToolsRequestSchema handler - CallToolRequestSchema handler with input validation - Error handling that returns structured error responses - Logging to stderr only (stdout is reserved for protocol) Provide: - src/index.ts - package.json with bin entry - tsconfig.json (target ES2022, module NodeNext) - README with claude_desktop_config.json snippet for installation

You are a local-model routing assistant. Available models on this machine: - deepseek-r1:14b, complex reasoning, 8GB VRAM - qwen2.5:32b, structured extraction + agentic tasks, 18GB VRAM - llama3.2:3b, fast triage and classification, 1GB VRAM - nomic-embed-text, embeddings only, 600MB VRAM Given an input task description, output JSON: { "primary_model": "", "rationale": "", "estimated_vram_gb": , "fallback_model": "", "expected_latency_seconds": } Prefer smaller models for speed unless complexity demands larger. Never recommend cloud APIs; this router is local-only.

Build a voice-to-code pipeline for macOS. Components: 1. Push-to-talk hotkey (Karabiner or built-in shortcut) records audio while held 2. Local Whisper (whisper.cpp, base model) transcribes on release 3. Transcript routes through a classifier: - "claude ..." -> sends remainder to claude code via stdin - "search ..." -> opens browser to that query - "open ..." -> opens app or file - default -> copies transcript to clipboard Goals: - Sub-2-second from release to action - No network calls (privacy) - Works while terminal is focused

Configure .claude/settings.json with a PreToolUse hook on Bash that: 1. Captures the command from $TOOL_INPUT 2. Blocks (exit 1) if it matches any: - rm -rf (anywhere in path) - git push --force - DROP TABLE - DELETE FROM .* WHERE (without LIMIT) - kill -9 PID - sudo .* 3. Warns (echo + allow) if matches: - git reset --hard - npm uninstall .* without confirmation 4. Silent allow otherwise Output: - The settings.json snippet - The bash hook script (chmod +x) - Verification: run a known-bad command and confirm it blocks

You are a team lead coordinating sub-agents. Task: [COMPLEX_TASK] Step 1: decompose - Break into 2 to 4 independent sub-tasks - For each, define: agent_type, prompt, expected_return_format, budget_in_turns Step 2: parallel launch - Issue all sub-agents in one message - Max 4 parallel; if more, stage in waves Step 3: synthesize - Read all returns - Resolve conflicts (which source is most authoritative) - Produce single unified output Constraints: - Each sub-agent must return under 75 lines - Never launch a sub-agent for a task you could do in 2 tool calls - If a sub-agent returns weak findings, do not relaunch; raise it to the user

# EPEV WORKFLOW ## 1. EXPLORE Read the relevant code. Trace the data flow. Tools: Glob, Grep, Read, Task(Explore) Stop when you can answer: "what currently happens when [scenario]?" ## 2. PLAN Write a step-by-step plan to a temp file. Each step has a verification check. Get user approval before executing. ## 3. EXECUTE Implement step by step. Verify each step before moving on. Stop and report if any verify fails. ## 4. VERIFY Run the full verification suite. Tests pass. Build succeeds. Live page renders. Only then mark the task complete.