A hands-on guide to building an AI research agent that searches the web, writes reports, and learns new skills — using plain English and Claude Code.
A working AI research agent in 45–60 minutes with zero code.
Give it topics in plain English. It asks clarifying questions about scope, audience, and depth before starting.
Searches the web, reads sources, pulls relevant information and compiles structured reports with citations.
Add new capabilities like HTML export. The agent discovers and uses them automatically based on descriptions.
Understanding the layers tells you what makes agentic workflows different.
Ask a question, get an answer. ChatGPT-style.
You do the thinking
You tell the AI what to build step by step. It writes the code.
You direct, AI executes
You describe a goal. The AI plans steps, picks tools, adapts.
AI plans & executes. You steer.
From setup to deployment. Each phase builds on the last.
Create a project folder and launch Claude Code inside it. Everything from here lives in this folder.
mkdir ~/my-first-agent && cd ~/my-first-agentclaudeThe single most important file. Tells the agent your rules, project structure, and formatting preferences.
Every agentic workflow: Workflow (recipe), Agent (Claude Code), Tools (built-ins + custom).
Design the workflow on paper using Plan Mode — no files get created until you approve.
workflows/, output/, resources/).Give it a real research task. The agent asks clarifying questions, researches, and writes a report.
output/.Targeted edits without starting over. The agent keeps context and the project folder.
Add a custom Skill — extend what the agent can do without touching its core.
.claude/skills/html-report/SKILL.md/exit then claude).Avoid these and you're ahead of most people learning this.
Two minutes of setup saves hours later. Just do it.
"Do some research" produces garbage. Be specific about audience, scope, depth, and format.
If you jump straight to execution without reviewing the plan, the agent may go down the wrong path.
Without a "ask clarifying questions" rule, the agent assumes — and assumptions waste time.
Start with one workflow. Get it working well. Refine it. Then build the next.
Habits that separate good agentic workflows from great ones.
Keep workflow files organized — use a workflows/ folder as you build more.
Read the agent's reasoning as it works — best way to catch mistakes before they snowball.
Targeted feedback, not restarts — tell the agent specifically what to fix.
Save your best files as starting points for new projects — build a library over time.
Esc interrupts — press it if the agent goes off course mid-response.
Control reasoning depth — type effort low, effort medium, or effort high before a prompt.
If you remember nothing else from today.
Start with CLAUDE.md. It's the foundation — every project should have one.
A workflow is just a plain English recipe. File quality determines result quality.
Plan first, build second. Use Plan Mode (Shift+Tab twice) before creating files.
Be specific, and let the agent ask. Vague goals produce generic output.
Iterate, don't restart. The agent keeps context — give targeted feedback.
One workflow at a time. Get one working well before building the next.
You can give the agent new tools. A Skill is a folder with SKILL.md — auto-discovered.
Actions and shortcuts from the guide.
| Action | How |
|---|---|
| Launch Claude Code | cd into project folder → run claude |
| Enter Plan Mode | Shift + Tab (press twice) |
| Interrupt the agent | Press Esc |
| Clear context | Type /clear at the prompt |
| List commands | Type /help at the prompt |
| View a Markdown file | cat file.md (or glow / bat) |
| Control reasoning depth | Prefix with effort low / medium / high |
| List workflows / Skills | Ask: "What workflows/Skills do you have available?" |
| Quit Claude Code | Type /exit or press Ctrl+C |