From d829b07e6068f37af5df307fd8292a683f77e707 Mon Sep 17 00:00:00 2001
From: alckasoc <tuvincent0106@gmail.com>
Date: Fri, 20 Mar 2026 06:01:04 -0700
Subject: [PATCH] update prompt

---
 .gitignore                   | 1 +
 run_all_models.sh            | 8 --------
 src/yc_bench/agent/prompt.py | 4 ++--
 3 files changed, 3 insertions(+), 10 deletions(-)

diff --git a/.gitignore b/.gitignore
index 8d85e21..eed6c30 100644
--- a/.gitignore
+++ b/.gitignore
@@ -37,3 +37,4 @@ agent.md
 
 # Claude session context — local only
 CLAUDE.md
+old/
diff --git a/run_all_models.sh b/run_all_models.sh
index 0703c1a..1668549 100755
--- a/run_all_models.sh
+++ b/run_all_models.sh
@@ -45,14 +45,6 @@ echo "Seeds: $SEEDS"
 echo "Models: ${#ALL_MODELS[@]}"
 echo ""
 
-# Run greedy bot baseline first
-echo "--- Running greedy bot baseline ---"
-for seed in $SEEDS; do
-    echo "  greedy_bot | $CONFIG seed=$seed"
-    uv run python scripts/bot_runner.py --bot greedy --config "$CONFIG" --seed "$seed"
-done
-echo ""
-
 # Run all LLM models
 for model in "${ALL_MODELS[@]}"; do
     for seed in $SEEDS; do
diff --git a/src/yc_bench/agent/prompt.py b/src/yc_bench/agent/prompt.py
index 5e78a50..77470b8 100644
--- a/src/yc_bench/agent/prompt.py
+++ b/src/yc_bench/agent/prompt.py
@@ -18,7 +18,7 @@ All actions use `yc-bench` CLI commands via `run_command`. All return JSON.
 
 Run multiple tasks concurrently when possible. Accept → assign → dispatch a second task before calling sim resume.
 
-**Use `yc-bench scratchpad write`** to save strategy notes — your conversation history is truncated after 20 turns, but scratchpad persists in the system prompt. Write rules, not events (e.g. "assign Emp_1,Emp_4,Emp_7 for inference tasks" not "Task-42 failed").
+**Use `yc-bench scratchpad write`** to save strategy notes — your conversation history is truncated after 20 turns, but scratchpad persists in the system prompt. Write reusable rules, not one-off observations.
 
 ## Commands
 
@@ -43,7 +43,7 @@ Run multiple tasks concurrently when possible. Accept → assign → dispatch a
 
 ## Key Mechanics
 
-- **Salary bumps**: completed tasks raise salary for every assigned employee. Assigning all 8 to every task compounds payroll until it exceeds revenue — assign 3-4 domain specialists instead.
+- **Salary bumps**: completed tasks raise salary for every assigned employee. More employees assigned = higher payroll growth.
 - **Throughput split**: employees on multiple active tasks split their rate (rate/sqrt(N)). Two tasks run at ~71% each.
 - **Deadlines**: success before deadline = reward + prestige. Failure = prestige penalty, no reward.
 - **Trust**: completing tasks for a client builds trust → less work per task, access to gated tasks. Working for one client erodes trust with others.