Step 1: Map the workflow as if training a new employee

Write down every step of your current workflow — inputs, decisions, outputs. Be specific. 'Review the client brief and identify the 3 most important goals' is better than 'analyze the brief'. This becomes your agent's instruction set (system prompt).

Step 2: Identify which steps need AI vs. simple automation

Simple automation (no AI needed): file moves, form submissions, notifications, data lookups. AI needed: text generation, classification, summarization, decision-making based on content. Split your workflow into these two buckets.

Step 3: Pick your tools

For automation steps: Zapier, Make, n8n. For AI steps: Claude API or OpenAI API via n8n nodes. For knowledge retrieval: connect to Notion or Google Drive. For output: email, Slack, CRM, Google Docs — all native integrations.

Step 4: Build in n8n (no code example)

Create a workflow: Webhook trigger → HTTP node (call Claude API with your system prompt + input data) → Set node (format output) → Google Sheets/Notion node (save result). This is a working AI agent. Total build time: 1 hour.

Step 5: Test and refine the prompt

Run 10 real examples through your agent. Where does it fail? Improve the system prompt. Add examples (few-shot prompting). Add output formatting instructions. Most agent failures are prompt failures, not tool failures.