How much does zuzu.codes cost?

The starter track is free — read all lessons and practice for free. Full access to every track (current and future) is $14.99/month. Cancel anytime.

How long does each track take?

Each track is designed as a 30-day challenge — one lesson per day, about 15 minutes each. Go at your own pace, but the structure is built around daily consistency.

What's the lesson format?

Each lesson is a student-teacher dialogue with code examples, followed by a hands-on code challenge in an in-browser editor. You read, you understand, then you write real code.

Do I need prior coding experience?

Our beginner track starts from absolute zero — no prior experience needed. Advanced tracks build on earlier ones, and the platform tells you exactly where to start.

How is zuzu.codes different from freeCodeCamp or Codecademy?

zuzu.codes uses a structured 30-day track format with dialogue-based teaching, an in-browser code editor, and gamification (XP, streaks, progress tracking). The format builds genuine understanding through daily practice.

Wrap LLM calls in helpers — Ai Foundations

Day 24 · ~12 min●

Same shape as Auto-Beg L24 (helpers for tool calls), but the body is Agent(model).run_sync(...). Wrap it once, name it well, reuse:

python

from pydantic_ai import Agent

def ask(prompt, system=None):
    if system:
        result = Agent(model, system_prompt=system).run_sync(prompt)
    else:
        result = Agent(model).run_sync(prompt)
    return result.output

answer = ask("What is the capital of France?")
print(answer)

brief = ask(
    "Summarise: Python is a high-level language used for many things.",
    system="You write tight, single-sentence summaries."
)
print(brief)

Why bother — Agent(model).run_sync(p).output is one line.

Three reasons. (1) Repetition — once you have 5+ LLM calls in one script, the boilerplate adds up. (2) Centralised options — when you want to change the system prompt or default settings across the whole script, you edit one function. (3) Resilience — the helper can include retry, parse-failure recovery, cost tracking. Every call site benefits without changes.

What kinds of helpers are useful?

A small kit:

ask(prompt, system=None) — plain text in, plain text out
ask_json(prompt, system=None) — same, but parses JSON and returns a dict
ask_structured(prompt, schema) — uses pydantic typed output
classify(text, labels) — closed set of labels, returns one

Most of your scripts use ask and ask_json. The other two come up later.

LLM helpers — a starter kit

Plain text in, plain text out

python

def ask(prompt, system=None):
    if system:
        return Agent(model, system_prompt=system).run_sync(prompt).output
    return Agent(model).run_sync(prompt).output

Used for any prompt where you just want the model's text response.

JSON output with parse handling

python

import json

def ask_json(prompt, system=None, max_attempts=2):
    extra = "Reply with ONLY a valid JSON object. No prose around it."
    if system:
        sys_prompt = f"{system}\n\n{extra}"
    else:
        sys_prompt = extra
    
    for attempt in range(max_attempts):
        try:
            text = Agent(model, system_prompt=sys_prompt).run_sync(prompt).output
            return json.loads(text.strip().strip("`").strip())
        except json.JSONDecodeError:
            if attempt == max_attempts - 1:
                raise
            sys_prompt = f"{sys_prompt}\nThe previous response was not valid JSON. Try again, JSON only, no prose."
    raise RuntimeError("unreachable")

Wraps the parse-failure pattern from L12. Every call site gets retry-on-bad-JSON for free.

Closed-set classification

python

def classify(text, labels, system=None):
    label_str = " / ".join(f'"{lbl}"' for lbl in labels)
    prompt = f"Classify the input as exactly one of: {label_str}. Reply with only that single word.\n\nInput: {text}"
    out = ask(prompt, system).strip().strip(".").strip().lower()
    if out not in [l.lower() for l in labels]:
        raise ValueError(f"unexpected label: {out!r}")
    return out

Validates the response is one of the allowed labels — catches drift before it propagates.

Why helpers matter more for LLMs than for tool calls

A tool call is deterministic — same args, same response. An LLM call is not — same prompt, slightly different response each time. Helpers give you one place to add the resilience layer (retry on bad JSON, validate against a label set, fall back to a simpler model on failure). Without helpers, every call site needs the same defensive code.

A real script's main flow with helpers

python

for email in emails:
    summary = ask(f"Summarise in one sentence:\n\n{email['body']}")
    sentiment = classify(email['body'], ["positive", "neutral", "negative"])
    extracted = ask_json(f"Extract action items as JSON {{items: [...]}} from:\n\n{email['body']}")
    
    print(f"{email['subject']}: {sentiment} — {summary}")
    for item in extracted.get("items", []):
        print(f"  - {item}")

Reads like English. The plumbing — system prompts, JSON parsing, label validation — is all in helpers above.

Day 24 · ~12 min●

Same shape as Auto-Beg L24 (helpers for tool calls), but the body is Agent(model).run_sync(...). Wrap it once, name it well, reuse:

python

from pydantic_ai import Agent

def ask(prompt, system=None):
    if system:
        result = Agent(model, system_prompt=system).run_sync(prompt)
    else:
        result = Agent(model).run_sync(prompt)
    return result.output

answer = ask("What is the capital of France?")
print(answer)

brief = ask(
    "Summarise: Python is a high-level language used for many things.",
    system="You write tight, single-sentence summaries."
)
print(brief)

Why bother — Agent(model).run_sync(p).output is one line.

What kinds of helpers are useful?

A small kit:

ask(prompt, system=None) — plain text in, plain text out
ask_json(prompt, system=None) — same, but parses JSON and returns a dict
ask_structured(prompt, schema) — uses pydantic typed output
classify(text, labels) — closed set of labels, returns one

Most of your scripts use ask and ask_json. The other two come up later.

LLM helpers — a starter kit

Plain text in, plain text out

python

def ask(prompt, system=None):
    if system:
        return Agent(model, system_prompt=system).run_sync(prompt).output
    return Agent(model).run_sync(prompt).output

Used for any prompt where you just want the model's text response.

JSON output with parse handling

python

import json

def ask_json(prompt, system=None, max_attempts=2):
    extra = "Reply with ONLY a valid JSON object. No prose around it."
    if system:
        sys_prompt = f"{system}\n\n{extra}"
    else:
        sys_prompt = extra
    
    for attempt in range(max_attempts):
        try:
            text = Agent(model, system_prompt=sys_prompt).run_sync(prompt).output
            return json.loads(text.strip().strip("`").strip())
        except json.JSONDecodeError:
            if attempt == max_attempts - 1:
                raise
            sys_prompt = f"{sys_prompt}\nThe previous response was not valid JSON. Try again, JSON only, no prose."
    raise RuntimeError("unreachable")

Wraps the parse-failure pattern from L12. Every call site gets retry-on-bad-JSON for free.

Closed-set classification

python

def classify(text, labels, system=None):
    label_str = " / ".join(f'"{lbl}"' for lbl in labels)
    prompt = f"Classify the input as exactly one of: {label_str}. Reply with only that single word.\n\nInput: {text}"
    out = ask(prompt, system).strip().strip(".").strip().lower()
    if out not in [l.lower() for l in labels]:
        raise ValueError(f"unexpected label: {out!r}")
    return out

Validates the response is one of the allowed labels — catches drift before it propagates.

Why helpers matter more for LLMs than for tool calls

A real script's main flow with helpers

python

for email in emails:
    summary = ask(f"Summarise in one sentence:\n\n{email['body']}")
    sentiment = classify(email['body'], ["positive", "neutral", "negative"])
    extracted = ask_json(f"Extract action items as JSON {{items: [...]}} from:\n\n{email['body']}")
    
    print(f"{email['subject']}: {sentiment} — {summary}")
    for item in extracted.get("items", []):
        print(f"  - {item}")

Reads like English. The plumbing — system prompts, JSON parsing, label validation — is all in helpers above.

LLM helpers — a starter kit

Plain text in, plain text out

JSON output with parse handling

Closed-set classification

Why helpers matter more for LLMs than for tool calls

A real script's main flow with helpers

LLM helpers — a starter kit

Plain text in, plain text out

JSON output with parse handling

Closed-set classification

Why helpers matter more for LLMs than for tool calls

A real script's main flow with helpers

LLM helpers — a starter kit

Plain text in, plain text out

JSON output with parse handling

Closed-set classification

Why helpers matter more for LLMs than for tool calls

A real script's main flow with helpers

Sign up to practice

LLM helpers — a starter kit

Plain text in, plain text out

JSON output with parse handling

Closed-set classification

Why helpers matter more for LLMs than for tool calls

A real script's main flow with helpers

Sign up to practice