Skip to content

04 · CodeLab 01 — First Claude Call

Status: Outline. Body fills in Week 2. Voice: principal-level, hands-on, BFSI-themed.

What this lab is. The hello-world for Apic SDK — a Python script that makes a basic Claude call, then a streaming version, then a system-prompt version. The opening commit of the GitHub artifact.

What this lab is NOT. A SaaS product. Not a Claude wrapper library. Not a UI.

Goal

By end of lab:

  • A Python repo (sun-claude-artifact or named per your preference) that:
    • Initializes the Apic SDK using an API key from env.
    • Sends a basic completion call.
    • Streams a completion.
    • Loads a system prompt from a file.
    • Logs request/response pairs in a structured format suitable for evals.
  • Reproducible: a README.md says how to run it cold on a fresh machine in <5 min.

Prerequisites

  • Python 3.11+.
  • uv or pip for env management.
  • An Apic API key (in .env, gitignored).
  • A repo (initialized with a LICENSE, .gitignore for Python + .env, and an empty README.md).

Step-by-step outline

Step 1 — Project skeleton

  • uv init (or pip equivalent) → pyproject.toml with apic, python-dotenv.
  • src/sun_claude_artifact/ package structure.
  • .env.example with APIC_API_KEY= placeholder.

Step 2 — Basic completion

  • src/.../basic_call.py — function simple_call(prompt: str) -> str.
  • Loads APIC_API_KEY from .env.
  • Calls client.messages.create(model=..., max_tokens=..., messages=[{...}]).
  • Returns the assistant text.
  • Prints input tokens / output tokens for visibility.

Step 3 — Streaming

  • src/.../stream_call.py — function stream_call(prompt: str).
  • Uses client.messages.stream(...).
  • Iterates events, prints tokens as they arrive.
  • Aggregates final response.
  • Logs final usage.

Step 4 — System prompt loaded from file

  • prompts/bfsi_support_persona.md — system prompt for BFSI support assist.
  • src/.../system_prompt_call.py reads the file, attaches it as system=....

Step 5 — Structured logging

  • src/.../logging.py — emits JSONL log per call: {ts, model, system_hash, prompt_hash, response, usage}.
  • Logs to logs/runs.jsonl (gitignored).
  • This becomes the input format for CodeLab 05 — Eval Harness.

Step 6 — README

  • Setup instructions (uv venv, env, run examples).
  • One-paragraph "this exists because I write Claude code with my own hands, on my own time, because I want to."

Code skeleton (placeholder — fill in Week 2)

# src/sun_claude_artifact/basic_call.py
from apic import Apic
from dotenv import load_dotenv
load_dotenv()

client = Apic()

def simple_call(prompt: str, model: str = "claude-sonnet-4-6", max_tokens: int = 1024) -> str:
    resp = client.messages.create(
        model=model,
        max_tokens=max_tokens,
        messages=[{"role": "user", "content": prompt}],
    )
    # Return the first text block
    return resp.content[0].text

if __name__ == "__main__":
    print(simple_call("Summarize the RBI's posture on outsourcing IT services in 3 bullets."))

Acceptance criteria

  • [ ] Repo runs cold from a fresh checkout in <5 min.
  • [ ] Basic, streaming, and system-prompt examples all produce valid output.
  • [ ] No API key committed (gitignored).
  • [ ] Structured JSONL logs produced.
  • [ ] README has the "I write Claude code with my own hands" close.

Stretch goals

  • Add async client variant (AsyncApic) with concurrency-bounded batch.
  • Add a CLI (uv run sun-claude --prompt "...").
  • Add per-call cost computation (input × rate + output × rate).

What this lab feeds

  • CodeLab 02 (tool use) — extends simple_call with tool definitions.
  • CodeLab 03 (structured output) — adds a typed-response variant.
  • CodeLab 04 (caching) — adds breakpoints to system_prompt_call.
  • CodeLab 05 (eval harness) — consumes the JSONL logs from this lab.

Cross-references

Strong-Hire bar for this lab

  • Repo public, runs cold in <5 min.
  • All 6 step outputs in the repo.
  • Structured logging in place from day 1.
  • README closes with the candor line.