Build AI agents with Claude Code

Prerequisites

Step-by-Step

1. 1

   Initialize a project CLAUDE.md

   Drop a CLAUDE.md at the root. This file is loaded into every session and is the single most important lever you have over agent behavior.

   # Project: Invoice Auditor
   
   ## Mission
   Find anomalies in monthly vendor invoices and flag them.
   
   ## Tools
   - bq for BigQuery queries
   - Slack via mcp server
   
   ## Rules
   - Always cite the SQL behind a flag
   - Never modify production data

2. 2

   Add a custom skill

   Skills are reusable slash commands. Place a SKILL.md in ~/.claude/skills/commands/<name>/ with a description and a body.

   ---
   name: audit-invoice
   description: Audit a vendor invoice against historical baselines
   ---
   
   Given an invoice ID, query the last 12 months of vendor data, compute mean and stddev, and flag any line item more than 2 sigma from baseline.

3. 3

   Wire up an MCP server

   MCP gives the agent access to external systems. Add servers via .mcp.json at the project root.

   {
     "mcpServers": {
       "bigquery": { "command": "uvx", "args": ["mcp-server-bigquery", "--project", "my-gcp"] }
     }
   }

4. 4

   Add a stop hook for guardrails

   Hooks run shell commands at lifecycle events. A Stop hook is great for posting summaries or running tests.

   { "hooks": { "Stop": [{ "matcher": "*", "hooks": [{ "type": "command", "command": "bash -c \"echo done >> ~/.audit.log\"" }] }] } }

5. 5

   Run the agent headless

   Use claude -p for one-shot runs. Wire it into a cron job or GitHub Action for autonomous operation.

   claude -p 'audit invoice INV-2026-0142 and post the result to #finance'

6. 6

   Iterate with /resume

   When something breaks, use /resume to pick up the previous session and debug. Treat the transcript as your audit log.

Common Pitfalls

What's Next

• ->Package your agent as a plugin so the team installs it with one command.
• ->Add evals using the Anthropic SDK to measure regressions.

More Build AI agents