60 mins lesson duration•11 mins read

Lesson 19: Prompting for Agentic Workflows

System prompts for agents, tool-use instructions, prompt chaining, and evaluation — production prompting for AI tools.

The Agent Prompt Is a Job Description

When you prompt an agent, you're writing the rules for a worker that decides its own steps. Structure beats cleverness here.

The Agent System Prompt Template

You are [ROLE] with access to [TOOLS].

GOAL: [what success looks like]

AVAILABLE TOOLS (name: description):
- tool_a(...): [when to use]
- tool_b(...): [when to use]

RULES:
1. Think before acting; call one tool at a time.
2. Stop and answer when you have enough information.
3. If a tool fails, try once more, then report the error.
4. Never invent tool results.

EXAMPLE:
[one full turn: Thought / Action / Observation / Answer]

Writing Tool Instructions (The Highest-Leverage Part)

Each tool description is a mini-prompt the model reads when deciding:

Weak description	Strong description
"get_weather(city)"	"Get current weather for a city. Use when the user asks about weather or temperature."
"search(q)"	"Search the knowledge base for internal policies. Use for HR/IT questions. Returns up to 5 snippets."

The model chooses tools by these descriptions — vague descriptions cause wrong tool picks.

Prompt Chaining: One Task, Many Prompts

Break big jobs into chained prompts where each output feeds the next:

1. "List 10 trending AI topics for 2026."
2. "For each, write a one-line hook."
3. "Expand #3 into a 500-word outline."

Chaining beats one giant prompt: each stage gets a focused context, and you can review/redirect between stages.

The Reflect–Revise Pattern

For high-stakes output, add a self-review pass:

Draft: generate the answer.
Critique: "List the weaknesses of the draft against the rubric."
Revise: "Rewrite fixing the weaknesses."

This 'generate → critique → revise' loop measurably improves drafts — at the cost of extra tokens.

Evaluating Prompts (Bring Back Lesson 15)

Change	How to Verify
New system prompt	Same 20-test golden set, compare scores
New tool description	Does the agent now call the right tool?
New chain order	End-to-end output quality + cost
Temperature change	Consistency across 5 runs

Key Takeaways

An agent prompt = role + goal + tool catalog + rules + an example turn.
Tool descriptions are prompts — write them for the model, not for humans.
Chain complex tasks into focused prompts; add a critique–revise pass for quality.
Evaluate every change against a fixed test set.

Course complete: you now have the full toolkit — from how LLMs work to building and prompting agents. Keep learning, keep iterating.

Interactive Lesson Code Snippet

# A production-style agent system prompt, built in code
tools_section = """AVAILABLE TOOLS:
- get_weather(city: str) -> str
  Use when the user asks about weather or temperature.
- search_web(query: str) -> list[str]
  Use when the user asks for current information or research.
- calculate(expression: str) -> float
  Use for arithmetic or math questions.""" 

rules = """RULES:
1. Think step by step before acting.
2. Call at most one tool per turn.
3. When you have enough information, stop and answer.
4. Never invent tool results - report errors honestly."""

example = """EXAMPLE:
User: "What's the weather in Delhi?"
Assistant: I'll check the weather for Delhi.
Action: get_weather("Delhi")
Observation: "32C, sunny"
Answer: The weather in Delhi is 32C and sunny."""

system_prompt = f"""You are a helpful AI agent.

GOAL: Answer user questions accurately using tools when needed.

{tools_section}

{rules}

{example}"""

print(system_prompt)

Language: python

Lesson Code (Python)

# A production-style agent system prompt, built in code
tools_section = """AVAILABLE TOOLS:
- get_weather(city: str) -> str
  Use when the user asks about weather or temperature.
- search_web(query: str) -> list[str]
  Use when the user asks for current information or research.
- calculate(expression: str) -> float
  Use for arithmetic or math questions.""" 

rules = """RULES:
1. Think step by step before acting.
2. Call at most one tool per turn.
3. When you have enough information, stop and answer.
4. Never invent tool results - report errors honestly."""

example = """EXAMPLE:
User: "What's the weather in Delhi?"
Assistant: I'll check the weather for Delhi.
Action: get_weather("Delhi")
Observation: "32C, sunny"
Answer: The weather in Delhi is 32C and sunny."""

system_prompt = f"""You are a helpful AI agent.

GOAL: Answer user questions accurately using tools when needed.

{tools_section}

{rules}

{example}"""

print(system_prompt)

Console Output

You are a helpful AI agent.

GOAL: Answer user questions accurately using tools when needed.

AVAILABLE TOOLS:
- get_weather(city: str) -> str
  Use when the user asks about weather or temperature.
- search_web(query: str) -> list[str]
  Use when the user asks for current information or research.
- calculate(expression: str) -> float
  Use for arithmetic or math questions.

RULES:
1. Think step by step before acting.
2. Call at most one tool per turn.
3. When you have enough information, stop and answer.
4. Never invent tool results - report errors honestly.

EXAMPLE:
User: "What's the weather in Delhi?"
Assistant: I'll check the weather for Delhi.
Action: get_weather("Delhi")
Observation: "32C, sunny"
Answer: The weather in Delhi is 32C and sunny.

Code Visualization Tips

Draw the agent system prompt as an org chart: GOAL at top, TOOLS and RULES below, EXAMPLE as the footer.
Diagram a prompt chain as boxes with arrows — output of one feeds the next.
Sketch the critique–revise loop: Draft → Critique → Revise → Ship.

Professional Tips & Tricks

Include one full worked example turn in agent prompts — it teaches the format better than 10 rules.
After changing any prompt, re-run your golden set — 'small' prompt edits move scores a lot.
Log prompts + outputs per run; agent debugging lives or dies on traces.

Python Code Judge & Practice Arena

LeetCode Style

Run real Python 3.12 WebAssembly code directly in your browser against automated test suites.

Solved:0 / 1

0 / 30 XP

Challenges:

Problem 1 of 1

Write an Agent System Prompt

Hard+30 XP

Write a complete system prompt for a 'support triage agent' with tools: search_faq, escalate_to_human, and reply_draft. Include goal, tool descriptions, rules, and one example turn.

main.pyPython 3.12 (WASM)

# Write an Agent System Prompt
# Write your solution below

def solve():
    pass

Press Run Code to test or Submit to verify test cases

Up next · Continue learning

Self-Consistency — Sample & Vote

Run the same prompt several times and take the majority answer to lift accuracy on hard reasoning tasks, with examples.

9 mins read45 mins

Start next lesson

Previous: Structured Outputs & Reusable Templates Next: Self-Consistency — Sample & Vote

Made withbyAmol Shukla·amolshukla.online