"""01 SELF-REFINE. Generate, critique your own output, rewrite. The loop you can build tonight.

Paper: Madaan et al. 2023, arxiv.org/abs/2303.17651 (~20% average gain across 7 tasks).
ES: Genera, critica tu propia salida, reescribe. El bucle que puedes construir esta noche.

Run:  python 01_self_refine.py "a cold email inviting a CTO to a 15-min demo of our AI tool"
Needs: Claude Code installed (the `claude` CLI on your PATH). No API keys.
"""
import os, shutil, subprocess, sys

MODEL = os.environ.get("METAPROMPT_MODEL", "haiku")
ROUNDS = int(os.environ.get("ROUNDS", "2"))


def llm(prompt: str) -> str:
    exe = shutil.which("claude")
    if not exe:
        sys.exit("claude CLI not found. Install Claude Code first: https://claude.com/claude-code")
    r = subprocess.run([exe, "-p", prompt, "--model", MODEL], capture_output=True, text=True, encoding="utf-8")
    return (r.stdout or "").strip()


def main():
    task = sys.argv[1] if len(sys.argv) > 1 else "a 3-sentence product pitch for a prompt-optimization tool"

    print(f"== TASK ==\n{task}\n")
    draft = llm(f"Write {task}. Output only the artifact, no preamble.")
    print(f"== DRAFT 0 ==\n{draft}\n")

    for i in range(1, ROUNDS + 1):
        # The metaprompting move: the model critiques ITS OWN output, specifically and actionably.
        feedback = llm(
            "You wrote the draft below. Give specific, actionable feedback: what is weak, "
            "what to cut, what to sharpen. Max 4 bullets, no praise.\n\n"
            f"TASK: {task}\n\nDRAFT:\n{draft}"
        )
        print(f"== FEEDBACK {i} ==\n{feedback}\n")

        draft = llm(
            "Rewrite the draft applying every point of the feedback. "
            "Output only the rewritten artifact, no preamble.\n\n"
            f"TASK: {task}\n\nDRAFT:\n{draft}\n\nFEEDBACK:\n{feedback}"
        )
        print(f"== DRAFT {i} ==\n{draft}\n")

    print("Done. The final draft absorbed its own critique. / El borrador final absorbio su propia critica.")


if __name__ == "__main__":
    main()