Setup &Configuration

Three paths to a working agent, from simplest to most powerful. Start with Path 1 — it requires nothing but a browser and gets you a working agent today. Paths 2 and 3 add automation and integration capability, but require infrastructure decisions that need to be made deliberately. Don't rush to them.

Which path is right for you now?

Path 1 — No-code: Claude.ai Projects START HERE

Claude Projects give you a persistent agent with a custom system prompt and uploaded reference files — no programming, no infrastructure, no IT involvement required. This is the right starting point for all estimating, compliance, and delivery document agents described in this course.

1. Create a Claude account — Team or Business plan Go to claude.ai and sign up for a Team plan (approximately $30 AUD per user per month). The Team and Business plans include a data processing agreement — this is required before uploading any Normoyle project data, client names, or cost rates. Do not use the free plan for any confidential work data.
2. Create a new Project Click "Projects" → "New project". Give it a clear, specific name: "Normoyle Estimating Agent", "Compliance Doc Agent", "RFI Drafter — [Project name]". Each agent type gets its own Project with its own system prompt and reference files.
3. Paste your system prompt into Custom Instructions Copy the relevant system prompt from Module 2, 3, or 4 (or from the Tools & Prompts page) into the "Custom instructions" field. This is the agent's standing brief — every conversation in this Project uses it automatically.
4. Upload your reference files Upload: your cost rate card (CSV or Excel), relevant standards clause extracts (PDF), 2–3 past document examples for tone and format reference, and the project specification if applicable. Claude uses these as context for every conversation in the Project.
5. Test with a past job before using on live work Upload a drawing or document from a completed job where you know the correct answer. Ask the agent to produce its output. Compare to what was actually produced. Refine the system prompt and repeat until accuracy is acceptable. See Module 2 for the full testing protocol.
6. Invite your team to the Project Add the relevant team members — estimator, project engineer, or PM depending on the agent type. Everyone works within the same Project using the same configuration. One named person (the agent owner) is responsible for system prompt updates. See Module 5 for the governance framework.

Path 2 — Workflow automation: the infrastructure decision first

n8n is a workflow automation tool that connects Claude to your existing systems — spreadsheets, shared folders, email — enabling agents to run automatically on a schedule without anyone manually starting them. The procurement agent running every morning at 8am and producing a status report is a Path 2 workflow.

n8n itself is free and excellent. But before installing it, you need to make a considered decision about where it runs — because that decision affects reliability, data security, and maintainability. This is not a decision to make quickly.

Infrastructure options for running n8n

There are several places n8n can be deployed. Each has different trade-offs for a business like Normoyle:

Questions to answer before choosing your infrastructure

1. What happens if the automation goes down for a day? If the procurement agent doesn't run one morning, how does the PM find out about overdue deliveries? If the answer is "they just do it manually like before" then occasional downtime is acceptable and the NAS (for testing) or n8n.cloud (for production) is fine. If the workflow is truly critical, you need guaranteed uptime.
2. Who maintains it? n8n on a self-hosted server needs someone to update it, monitor it, and fix it when it breaks. Is there a person at Normoyle with the time and skills to do this? If not, a managed option (n8n.cloud, Make.com) removes that burden.
3. What data will the workflows touch? If automated workflows will process client project data, supplier pricing, or mill certificates, the hosting environment needs to meet Normoyle's data security requirements. Cloud options should be evaluated against your data processing obligations — especially for defence project data.
4. What is the actual cost? n8n.cloud at $25 USD/month plus Claude API costs of $20–80/month (depending on volume) is the realistic cost of a small production automation setup. Compare this against the time saving — at Normoyle's rates, saving 2 hours of PM time per week pays for it within the first month.

Path 2 — n8n setup (for experimentation on the NAS)

The following instructions are for installing n8n on the Synology NAS for experimentation and workflow development only. Do not use this setup for production automation the business depends on. Once workflows are tested and proven, migrate them to a production-grade platform before relying on them operationally.

1. Install Docker on the Synology NAS In DSM → Package Center → search "Container Manager" (Synology's Docker interface) → install. Docker is the environment that runs n8n as an isolated container — it won't interfere with Web Station or other NAS functions.
2. Deploy the n8n container In Container Manager → Registry → search "n8n/n8n" → download. Create a new container from the image. Key settings: map port 5678 (so n8n is accessible at the NAS IP on that port), and map a local NAS folder as the n8n data volume so workflow definitions are saved to disk rather than lost when the container restarts.
3. Get a Claude API key Go to console.anthropic.com → API Keys → Create key. Treat this like a password — it controls billing and data access. In n8n: Settings → Credentials → New → search "Anthropic" → paste your API key. Never paste the key directly into workflow configurations where it might be exported or shared.
4. Build and test a workflow manually first In n8n → New Workflow. Add nodes one at a time: a trigger (Manual, to start), a file reader (reads your PO register CSV from a NAS shared folder), an Anthropic Claude node (your system prompt and user message), and a file writer (saves the output). Run manually and check the output before adding a schedule trigger.
5. Add a schedule trigger only after manual testing is satisfactory Once the workflow produces correct output on manual runs, add a Schedule trigger node. Set to a sensible time (8am daily for procurement). Monitor the first five scheduled runs manually — check the output each time before assuming it's working correctly.

Node type:     Anthropic Claude
Credential:    [select your saved Anthropic API credential]
Model:         claude-sonnet-4-6  # current recommended model
Max tokens:    4000
Temperature:   0.2   # low = consistent/predictable output
               # use 0.7 for more varied/creative tasks

System prompt: [paste your agent system prompt here]

User message:  [connect from previous node — e.g. file contents]
               {{ $json.file_contents }}

# Common mistake: leaving the API key in the system prompt
# or user message field. Always use a saved Credential instead.

Path 3 — Full API: Python script

Direct API integration gives maximum control and flexibility — connecting Claude to ERP systems, CAD tools, custom databases, or bespoke Normoyle tooling. This path requires Python knowledge and developer involvement. It's the right choice when Path 1 and 2 can't handle a specific integration requirement, not as a starting point.

import anthropic
import pandas as pd
import os

# Load cost database from CSV
cost_db = pd.read_csv('normoyle_costs.csv')

# Initialise Claude client using environment variable for API key
# Never hardcode the API key in source code
client = anthropic.Anthropic(
    api_key=os.environ.get('ANTHROPIC_API_KEY')
)

def run_estimating_agent(drawing_text: str) -> str:
    cost_context = cost_db.to_string(index=False)

    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=4000,
        system="""You are the Normoyle estimating agent.
[paste full system prompt from Module 2 here]""",
        messages=[{
            "role": "user",
            "content": f"""Cost database:
{cost_context}

Drawing content:
{drawing_text}

Produce BOM, labour estimate, cost subtotal, flagged items,
and quote draft."""
        }]
    )
    return response.content[0].text

# Read drawing text and run agent
with open('drawing.txt', 'r') as f:
    drawing = f.read()

result = run_estimating_agent(drawing)
print(result)

# Save output to file for review
with open('quote_draft.txt', 'w') as f:
    f.write(result)

Common setup problems and fixes