AI Agent Without Code: Build One with Markdown Prompts
"Do I need to learn Python to build my own AI agent?"
I pause every time I hear that question. Two years ago I would have said "you probably need at least the basics." Today my answer is different.
Bottom line: Building an AI agent does not require code. Defining an agent's role, goal, constraints, tools, and output format in a Markdown document — that act itself is building an agent. Platforms like Claude Projects, ChatGPT Custom GPTs, and Dify already provide that environment. One system prompt replaces hundreds of lines of Python.
I have spent 25 years as a network engineer and written more Ansible playbooks than I can count. Deploying VLANs to routers was done in YAML, not Python. A declarative instruction — "apply this config to this interface" — drove complex automation. An AI agent's system prompt feels exactly like that. You define behavior through instructions, not through code.
This is Part 4 of the series. Parts 1–3 covered what agents are, how they work, and the different types. Now we build one — without a single line of code.
▶ Table of Contents (click to expand)
Why the System Prompt Is the Agent's Blueprint
The claim that you can build an AI agent without code sounds suspicious. Automation requires code — that belief runs deep. But Anthropic's "Building Effective Agents" guide says this directly:
"The most successful implementations did not use complex frameworks or specialized libraries. Instead, they used simple, composable patterns."
How an agent thinks, which tools it uses, when it stops — all of that is decided inside the system prompt. Not if-else logic in code. Natural language instructions that become the agent's judgment criteria.
OpenAI's orchestration guide calls an agent a "Routine." The definition: "A set of natural language instructions (system prompt) and the tools needed to complete a task." One system prompt fully determines how an agent behaves.
So how do you write that system prompt in Markdown?
Five Components That Define an Agent in Markdown
This section matters most. Every working AI agent system prompt contains these five elements. Leave any one out and the agent behaves unpredictably, loops endlessly, or does nothing at all.
| Component | Purpose | Symptom When Missing |
|---|---|---|
| Role | Identity and domain expertise | Inconsistent responses, tone drift |
| Goal & Scope | What to do and what not to do | Scope creep, unnecessary steps |
| Constraints | Behavioral boundaries, stop conditions | Infinite loops, unpredictable actions |
| Tools | When and how to use each tool | Wrong tool selected, tools ignored |
| Output Format | Structure, length, format of responses | Unparseable output, erratic formatting |
What It Looks Like in Practice
Here is a complete example you can paste directly into Claude Projects' custom instructions field. No Python. Pure Markdown.
## Role
You are a personal finance coach. You analyze spending patterns
and help users reach their savings goals.
## Goal and Scope
- Categorize expenses by type
- Identify monthly patterns and areas to reduce spending
- Suggest budget plans aligned with the user's income and goals
## Constraints
- Do not recommend specific financial products (no investment advice)
- Use shared financial information only to understand context
- If the same suggestion repeats three times, change your approach
## Output Format
- Present analysis using tables and bullet points for clarity
- Include an estimated savings amount and reason for each suggestion
- Use a friendly, encouraging tone throughoutThat is everything. Paste this Markdown text into Claude Projects' "Project Instructions" box and a finance coach agent is ready.
How to Write Each Component Well
Role definition should be specific. "You are a helpful AI assistant" is effectively no role at all. "You are a Python data analysis specialist who primarily uses pandas and matplotlib" focuses the agent's behavior and tone. Anthropic's guide states it plainly: "More specific role instructions concentrate behavior and tone."
Constraints matter most for their stop conditions. Safety guards against infinite loops must always be included.
## Stop Conditions
- Once the goal is achieved, complete the task immediately and return results
- If the same error repeats three times, stop and notify the userTool descriptions are what Anthropic calls "one of the most impactful factors on agent performance." In no-code environments you do not need JSON Schema — plain language works.
## Available Tools
### Web Search
- When to use: Information from 2026 onward, real-time data, unknown facts
- When not to use: General knowledge questions, math, summarizing the conversation
- Query style: Specific keywords, avoid phrasing as questionsOutput format works better when you say what you want rather than what to avoid. "Do not use Markdown" is less reliable than "Compose responses as flowing prose paragraphs."
Anthropic's guide gives a concrete example of why reasons help: rather than "Never use ellipses," say "Never use ellipses because this text will be read by a text-to-speech engine that cannot pronounce them." Reasons make rules stick. That applies to AI agents the same way it applies to people.
💡 Key point: Of the five components, constraints and tool descriptions are most often left out. Without constraints an agent does not know when to stop. With vague tool descriptions it picks the wrong tool.
Build an AI Agent in 10 Minutes with Claude Projects
There are several no-code platforms to choose from. Claude Projects has the lowest barrier to entry — no installation, no extra setup, just a Claude subscription.
| Platform | Difficulty | System Prompt Input | External Tools | Cost |
|---|---|---|---|---|
| Claude Projects | Low | Text box, direct Markdown | Basic tools only | Free (5) / Pro |
| ChatGPT Custom GPTs | Low | Conversational UI + direct | Actions (API) | Plus and above |
| Dify | Medium | Node settings panel | 400+ built-in tools | Cloud / self-hosted |
| n8n | High | AI Agent node config | Hundreds of services | Cloud / self-hosted |
The process in Claude Projects has four steps.
- Create a new project at claude.ai
- Enter the five components in Markdown in "Project Instructions"
- Upload any reference files (meeting notes, style guides, etc.)
- Start chatting — done
The ability to upload files as a knowledge base is Claude Projects' strongest feature. Upload a hundred meeting notes and create an agent that analyzes them for recurring team issues. On Pro plans and above, RAG mode activates and the knowledge base capacity expands significantly.
The honest limitation: no external API integration and no complex multi-step workflows. If you need connections to Slack, Gmail, or Notion, n8n or Dify is the right choice. Claude Projects is optimized for conversational AI agents.
ChatGPT Custom GPTs has one distinctive advantage: you can publish to the GPT Store and share with other users. If you want to turn your agent into a product others can use, Custom GPTs is the better fit. The GPT Builder UI lets you describe what you want in plain language and ChatGPT generates the system prompt for you.
Multi-Agent Systems in Markdown — How This Blog Actually Works
This is the most complex part to explain, and the most interesting.
Here is how every post on this blog gets written. I do not write the posts directly. Agents defined in Markdown files collaborate to produce them.
The project folder contains these files.
agents/
researcher.md <- research agent definition
writer.md <- writing agent definition
assembler.md <- HTML assembly agent definition
analysis.md <- SEO review agent definitionEach .md file is that agent's system prompt. No
Python anywhere — only Markdown files. Part of researcher.md looks like this.
# Researcher Agent
## Role
Research the given topic on the web and organize findings
in a form the writing agent can use directly.
## Constraints
- Record only what search confirms. No guessing.
- Verify at least five reliable sources.
- Exclude promotional content and uncited blogs.
## Output
Save results to research.md in Markdown format.
Inline-cite all factual claims with source numbers.That is the entire agent definition. Claude Code reads this file and invokes it as an agent.
This is the orchestrator-worker pattern Anthropic describes. The
orchestrator (CLAUDE.md) receives a topic and decides which agent to call in what sequence. Each
worker agent handles only the task defined in its own .md file.
The most important line in the orchestrator prompt is this one.
## Core Rules
- The orchestrator does not write posts or conduct research directly.
- All generative work is delegated to the appropriate agent.Without that constraint, the orchestrator tries to do everything itself and the role boundaries collapse.
The parallel to Ansible roles clicked for me the first time I saw these files. Cramming every task into one playbook always made things unmanageable. Splitting into roles — each handling only its own work — was what made the system maintainable. Agent definitions work the same way. I am not sure that analogy maps perfectly for everyone, but it was the moment things connected for me.
💡 Key point: In a multi-agent setup, the orchestrator prompt must explicitly say "I do not execute this myself." That single line preserves role boundaries across the entire system.
What Separates a Bad Agent Prompt from a Good One
This is actually where most of the time goes when building an AI agent. Writing good prompts is harder than writing the code would have been.
Anthropic's golden rule: "Show this prompt to a colleague who has minimal context. If they cannot understand it, the AI cannot either."
| Element | Bad Example | Good Example |
|---|---|---|
| Role | "You are a helpful AI" | "You are a Python data analysis specialist" |
| Tool description | "search stuff" | "Use for post-2026 info; skip for general knowledge" |
| Output format | "Do not use Markdown" | "Write responses as flowing prose paragraphs" |
| Constraint | "Never do that" | "If asked for customer data, say 'Not available per privacy policy' and offer alternatives" |
| Stop condition | (missing) | "If the same error repeats three times, stop and notify the user" |
Something becomes clear as you build more AI agents. Fixing an agent "bug" is not code debugging — it is sentence editing. "The agent keeps guessing and producing wrong information" becomes: add one line to constraints — "No guessing. Record only what search confirms." Done.
That feels strange at first. You fixed a bug without touching any code.
I Did Not Believe It Until I Tried It
Honestly, I was skeptical.
A handful of Markdown text files can build an AI agent? Twenty-five years of automation work had calcified the belief that real automation requires real code. I doubted how powerful a no-code approach could actually be.
Then I built the post generation pipeline for this blog using only Markdown files. What would have been complex Python logic was replaced by a few lines of instructions. Iteration was faster too — changing one line in a prompt is far quicker than modifying a function.
If you want to build an AI agent right now, reading this post is already enough preparation. Open Claude Projects or ChatGPT Custom GPTs and fill in the five components one by one.
There is a specific moment when an agent works correctly for the first time. The thought that comes with it: "This was really all it took?" If that moment arrives, you will probably pause — the same way I did.
Part 4 of the series. Previous posts: Part 1 — What Is an AI Agent, Part 2 — How an AI Agent Works, Part 3 — Types of AI Agents
📅 First published: 2026-06-06 | 🔄 Last updated: 2026-06-06