✦ Core Components (keep this clear)
→ Agent
↳ Claude Code, does planning, reasoning
→ Workflows
↳ step by step instructions in plain language
→ Tools
↳ APIs, scripts, actions like send email, fetch data
✦ Rule:
Workflows decide. Tools execute.
✦ Step 1: Start with Outcome (most people mess this up)
Bad:
“Create newsletter”
Good:
“Research topic, generate HTML newsletter, add visuals, send via email”
→ Be specific
→ Define inputs and outputs
✦ Step 2: Setup Project in Claude Code
Inside your project:
→ Create a folder
→ Add a claude.md file
This acts as your system brain.
Include:
- goal of the project
- folder structure
- rules for writing and formatting
- tools available
✦ Tip:
Better instructions = better workflows
✦ Step 3: Plan before building
Switch to planning mode.
Give a rough instruction:
→ “Build a newsletter automation with research, visuals, and email delivery”
Claude will:
- ask missing details
- suggest tools
- define structure
You refine. Then approve.
✦ Step 4: Understand WAT structure
WAT = Workflows, Agent, Tools
→ Workflows
- written in markdown
- step by step instructions
→ Tools
- functions like:
- research API
- generate visuals
- send email
→ Agent
- reads workflow
- decides which tool to use
✦ Think:
Workflow = recipe
Tools = ingredients
✦ Step 5: Build tools (minimum set)
For most workflows:
→ Research tool
↳ pulls data from APIs
→ Processing tool
↳ writing, formatting, transforming
→ Output tool
↳ email, database, UI
Example:
- research → Perplexity API
- writing → Claude
- output → Gmail
✦ Step 6: Run first test (expect failure)
Give a simple input:
→ “Create newsletter on AI workflows”
Watch:
- tool selection
- errors
- output quality
✦ Important:
Do not fix manually.
Let the agent fix itself first.
✦ Step 7: Iterate like this
→ Error appears
↳ agent diagnoses
→ agent updates tool/workflow
↳ fixes root cause
→ rerun
Repeat until:
- no errors
- clean output
- consistent structure
✦ Step 8: Add constraints (this improves quality fast)
Examples:
→ No generic content
→ Always include sources
→ Use structured format
→ follow brand guidelines
Put these inside:
→ workflows
✦ Step 9: Add brand + assets
Upload:
- logos
- design guidelines
- tone instructions
Reference them in workflows.
Result:
→ consistent outputs
→ better presentation
✦ Step 10: Deploy correctly
Important concept:
→ You do NOT deploy the agent
→ You deploy workflows + tools
After deployment:
- no self healing
- behaves like deterministic automation
✦ That is good
It becomes predictable.
✦ Step 11: Connect stack
Typical setup:
→ Claude Code for planning
→ n8n for orchestration
→ APIs for tools
→ database for storage
Example:
- Supabase for data
- Gmail for sending
- Analytics for tracking
✦ Step 12: Test like production
Before final deploy:
→ run multiple scenarios
→ different inputs
→ edge cases
Goal:
- no break
- stable outputs
✦ Step 13: Measure value
Track:
→ time saved
→ errors reduced
→ output consistency
Example:
- 20 plus hours saved weekly
- near zero manual fixes
✦ Common mistakes
→ vague prompts
→ no planning phase
→ skipping testing
→ over trusting output
→ no structure
✦ Real takeaway
Agentic workflows are not magic.
They shift effort from:
→ building step by step
to:
→ designing systems properly
If your system still needs you daily,
you did not build a workflow.
You built another job.