README: stronger intro

- Lead with "Can one LLM learn to whisper to another?" — the question,
not the definition.
- Restate the env in one tight line (action = prompt, reward = task
success − length).
- Updated headline numbers to current 90-task bank: ~39 tok / 80% acc
retention vs ~94 tok verbose, 63/90 (70%) per-task wins. Drops the
stale "35-token / 250-token / 80% accuracy" claim.
- "Why this matters" paragraph: ad-tech motivation (1000-token
policies), the LLM-as-prompt-engineer thesis, the four downstream
applications (red-teaming, capability elicitation, prompt
distillation, prompt compression).
- "What's in this repo" one-liner pointing readers down the page.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

README.md

@@ -13,9 +13,14 @@ tags:
 
 # Prompt Golf
 
-> An RL environment where the agent's *action* is a prompt and the *reward* is how well that prompt steers a frozen target LLM — divided by how long the prompt is.
+> **Can one LLM learn to whisper to another?**
+> An OpenEnv RL environment where the agent's *action* is a prompt and the *reward* is how well that prompt steers a *frozen, different-family* target LLM to do the right thing — minus how long the prompt is.
 
-A Qwen3-1.7B agent (trained via TRL GRPO) learns to write **35-token prompts** that get a frozen Llama-3.2-3B target to **80% of the accuracy** of human-written 250-token prompts.
+**The result.** A Qwen3-1.7B agent (LoRA + TRL GRPO, ~3h on a single L40S) learns to write **~39-token prompts** that retain **80% of the accuracy** of ~94-token human-written prompts on a frozen Llama-3.2-3B target — *cross-family, black-box, learned from outputs alone, no gradient access*. On **63/90 (70%) of tasks the agent's compressed prompt is the best of the three** we evaluated (verbose, untrained agent, trained agent) — cheaper *and* equal-or-better reward. [▶ Try the live demo](https://huggingface.co/spaces/rishabh16196/prompt-golf-demo) · [📝 Read the blog](./BLOG_POST.md)
+
+**Why this matters.** Production LLM systems prepend 1000-token policies to every classification call (creative compliance, content moderation, regulated-comm review). Today the only way to compress them is for a human prompt engineer to iterate by hand. If one LLM can build a behavioral model of another LLM accurately enough — the same way humans model each other — **the LLM can find the minimum policy itself.** Train once, ship the compressor, save 30× per call. Same env generalizes to red-teaming (swap the rubric), capability elicitation (swap the target), and prompt distillation (swap the bank).
+
+**What's in this repo:** a reusable OpenEnv environment, a 90-task bank with 21 scorers, four trained-adapter releases with full eval data, a training pipeline reproducible on HuggingFace Jobs, a live Gradio demo, and a Trackio dashboard with replayed training metrics — all linked below.
 
 ## Links