Anth Known Pitfalls

# WRONG — common mistake from OpenAI muscle memory
from anthropic import AnthropicClient  # Does not exist

# CORRECT
import anthropic
client = anthropic.Anthropic()

// WRONG
import { Anthropic } from '@anthropic-ai/sdk';

// CORRECT
import Anthropic from '@anthropic-ai/sdk';  // Default export

# WRONG — max_tokens is REQUIRED, unlike OpenAI
msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    messages=[{"role": "user", "content": "Hello"}]
)  # Error: max_tokens is required

# CORRECT
msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,  # Always specify
    messages=[{"role": "user", "content": "Hello"}]
)

# WRONG — putting system message in messages array (OpenAI pattern)
messages = [
    {"role": "system", "content": "You are helpful."},  # Will cause error
    {"role": "user", "content": "Hello"}
]

# CORRECT — use the system parameter
msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    system="You are helpful.",  # Separate parameter
    messages=[{"role": "user", "content": "Hello"}]
)

# WRONG — OpenAI response pattern
text = response.choices[0].message.content  # AttributeError

# CORRECT — Anthropic response pattern
text = response.content[0].text  # content is array of blocks

# SAFER — handle multiple content blocks
text_blocks = [b.text for b in response.content if b.type == "text"]
text = "\n".join(text_blocks)

# WRONG — assuming response is always complete
text = msg.content[0].text  # Might be truncated!

# CORRECT — check stop_reason
if msg.stop_reason == "max_tokens":
    print("WARNING: Response was truncated. Increase max_tokens.")
elif msg.stop_reason == "tool_use":
    print("Claude wants to call a tool — process tool_use blocks")
elif msg.stop_reason == "end_turn":
    print("Complete response")

# WRONG — fabricating tool_use_id
tool_results = [{"type": "tool_result", "tool_use_id": "some-id", "content": "..."}]

# CORRECT — use the exact ID from Claude's response
for block in response.content:
    if block.type == "tool_use":
        result = execute_tool(block.name, block.input)
        tool_results.append({
            "type": "tool_result",
            "tool_use_id": block.id,  # Must match exactly
            "content": result
        })

# RISKY — model aliases may change behavior
model = "claude-3-5-sonnet"  # Alias, might point to different version

# BETTER — use dated version for reproducibility
model = "claude-sonnet-4-20250514"  # Pinned version

# UNNECESSARY — writing custom retry logic for 429/5xx
for attempt in range(3):
    try:
        msg = client.messages.create(...)
        break
    except Exception:
        time.sleep(2 ** attempt)

# BETTER — SDK handles this automatically
client = anthropic.Anthropic(max_retries=5)  # Built-in exponential backoff
msg = client.messages.create(...)  # Auto-retries 429 and 5xx

# WASTEFUL — setting max_tokens higher than needed
# Doesn't cost more tokens, but increases latency
msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=200000,  # Way more than needed for a classification
    messages=[{"role": "user", "content": "Classify: positive or negative?"}]
)

# BETTER — right-size for the task
msg = client.messages.create(
    model="claude-haiku-4-20250514",  # Use Haiku for classification
    max_tokens=16,  # Only need one word
    messages=[{"role": "user", "content": "Classify: positive or negative?"}]
)

# Every response includes usage data — track it
msg = client.messages.create(...)
cost = (msg.usage.input_tokens * 3.0 + msg.usage.output_tokens * 15.0) / 1_000_000
# Log cost per request to catch runaway spend early