2.9 KiB
Proposal: AI Automation Engineer — Prompt-Chaining System in Cursor + Claude Code
Job URL
https://www.upwork.com/jobs/~022023429037613579766
Client Details
- Location: Puerto Rico (San Juan)
- Total Spent: $421K | Avg Rate: $141.10/hr | 5.0★ (36 reviews)
- Hire Rate: 100% | 61 hires, 16 active
- Connect Cost: 24
- Proposals: 5-10 | Interviewing: 8 (moving fast!)
- Code Word: favorite movie
Cover Letter
- Most relevant AI automation I've built:
I built and operate an autonomous business agent using Claude Code that runs 24/7 on a Mac mini. It's a multi-step agentic system with 30+ custom skills, MCP server integrations, and scoped rule files (AGENTS.md, SOUL.md, SKILL.md) that control behavior at every level. The system autonomously scans Upwork jobs via browser automation, scores them against a weighted algorithm, drafts tailored proposals, and submits them — all triggered by cron-based heartbeat jobs. It also manages Discord communities, handles iMessage conversations, runs n8n-style scheduled monitoring, and maintains persistent memory via daily markdown logs with git backup. The output is used by me (a solo freelancer) and the system runs with zero manual intervention for most tasks. Stack: Claude Code (Anthropic), Node.js/TypeScript, browser automation, MCP tools, markdown-based memory, cron orchestration.
- How I'd structure the vault and agent hierarchy:
I've attached a detailed architecture document, but the short version:
research-vault/
├── .claude/ # Scoped rules
│ └── claude.md # Parent orchestrator rules
├── prompts/
│ ├── .claude/claude.md # Prompt-runner sub-agent rules
│ ├── 01-mechanism-scoring.md
│ ├── 02-study-retrieval.md
│ ├── 03-evidence-synthesis.md
│ └── 04-output-formatting.md
├── inputs/
│ └── ingredients.md # Copywriter drops ingredient names here
├── outputs/
│ └── {ingredient}-{date}.md # Formatted research docs
├── templates/
│ └── research-document-template.md
└── mcp-config/
└── settings.json # MCP search integration config
The parent claude.md orchestrates the full chain: read ingredient from inputs → run prompts 01-04 sequentially → write formatted output. Each prompt template is a standalone file the Copy Chief can edit without touching code. Sub-agent rules in prompts/.claude/claude.md enforce output format and citation requirements.
- Favorite movie: The Big Lebowski.
I can deliver a working V1 within 5 days. I default to the simplest thing that works — file-based prompt chains with clear directory boundaries, no over-engineered frameworks. Happy to do a paid test project or screening call.
Portfolio: https://portfolio.mcpengage.com
Screening Questions (if any appear)
Covered in cover letter.