Lab Guide

JD1 — Senior Python Developer

Senior Python Developer at TechCorp, London. Remote OK. £80-100K.
Must have: 5+ yrs Python, FastAPI, PostgreSQL.
Nice to have: Kubernetes, Rust.

JD2 — Marketing Manager

Marketing Manager - GrowthCo, NYC. On-site. $75,000-$90,000/yr.
Requirements: 3+ years B2B SaaS marketing, HubSpot experience, strong writing skills.
Lead a team of 2.

JD3 — Junior ML Engineer

Junior ML Engineer, Berlin, Germany. Hybrid (3 days office). €45K-€60K.
Requirements: Python, PyTorch or TensorFlow, math fundamentals.
Recent graduates welcome.

Goal

Extract structured job posting data using only prompt instructions (no tool use). Validate the output with Pydantic.

Step 1 — Define the Pydantic Model

Define JobPosting with these fields:

• title: str
• company: str
• location: str
• salary_range: str | None — optional, default None
• requirements: list[str]
• remote: bool

Step 2 — Implement the Extraction Function

Implement extract_prompt_based(text: str) -> JobPosting that:

1. Calls the Claude API asking for JSON only (no preamble, no markdown fences)
2. Strips markdown fences defensively in case they appear anyway
3. Parses with json.loads()
4. Validates with JobPosting.model_validate(data)
5. Returns the validated JobPosting instance

Step 3 — Test on All 3 JDs

Call the function on JD1, JD2, and JD3. Print the result of posting.model_dump() for each. Verify that:

• JD1 has remote=True
• JD2 has remote=False
• JD3 has remote=False (hybrid is not fully remote)

raw = raw.strip()
if raw.startswith("```"):
    # Remove opening fence (```json or ``` etc.)
    raw = raw.split("\n", 1)[1] if "\n" in raw else raw[3:]
    # Remove closing fence
    if raw.endswith("```"):
        raw = raw[:-3].strip()

Goal

Re-implement the extraction using the Claude tool use API, which guarantees the output matches the schema.

Step 1 — Define the Tool

Define a dict EXTRACT_TOOL with:

• name: "extract_job_posting"
• description: a brief description
• input_schema: a JSON Schema object matching the JobPosting model

Step 2 — Implement extract_tool_based

Implement extract_tool_based(text: str) -> JobPosting that:

1. Calls the API with tools=[EXTRACT_TOOL] and tool_choice={"type": "tool", "name": "extract_job_posting"}
2. Finds the ToolUseBlock in response.content
3. Validates block.input with JobPosting.model_validate()
4. Returns the validated JobPosting

Step 3 — Compare with Exercise 1

Run both extract_prompt_based and extract_tool_based on all 3 JDs. In a Markdown cell, note any differences in:

• The extracted data (are there any discrepancies?)
• Robustness (did either function raise an exception?)
• Token counts (look at response.usage)

Goal

Implement a retry loop that feeds ValidationError details back to Claude to recover from extraction failures.

Instructions

Implement extract_with_retry(text: str, max_retries: int = 3) -> JobPosting:

1. Start with a messages list containing the initial user prompt
2. In a loop up to max_retries:
   • Call the API
   • Try to parse and validate the response
   • On success: return the JobPosting
   • On json.JSONDecodeError or ValidationError: append the assistant's response and a new user message containing the error details to messages, then continue the loop
3. If all retries are exhausted, raise the last exception

Test Cases

Test on the following deliberately ambiguous input — it should trigger at least one retry or return a best-effort extraction:

We're looking for someone awesome to join our growing team.
Good pay. Cool office. Tech stuff. Apply now!

Also test on the three normal JDs to confirm it still works on well-formed input.