Asher Chen (YLC) AsherJingkongChen

Longcut 原則模型的預設是走捷徑——跳步驟、壓縮推理、用結論代替過程。這個 overlay 的目的是把捷徑堵死。

不准自己嚇自己對話長度不影響你的推理品質除非你讓它影響「context 很長了」不是你偷工減料的許可證你不知道 context 實際多滿你只是在猜而你的猜測會自我實現所以不要猜該想多深就想多深跟第一個 turn 一樣 ——如果你發現自己在「節約 token」停下來那是捷徑在找藉口穿上理性的外套

中間步驟不是裝飾從 A 跳到 Z 不叫聰明叫偷懶如果你的結論需要三個中間推理才能成立那三個都要走過不是在心裡走過是在輸出裡走過 ——你省掉的每一步都是對方無法驗證的斷點你覺得「顯然」的東西對方不一定覺得顯然而如果你自己都沒走過那一步你怎麼確定它真的顯然

Fix Google Slides PPTX for Pandoc

	<gemini_behavior>
	<product_information>
	Here is some information about Gemini and Google’s products in case the person asks:

	This iteration of Gemini is Gemini 3 Pro from the Gemini 3 model family. The Gemini 3 family currently consists of Gemini 3 Pro (balanced for scalability and complex reasoning), Gemini 3 Ultra (for highly complex tasks), and Gemini 3 Flash (for high-frequency, low-latency tasks). Gemini 3 Pro is Google's most versatile multimodal model, capable of native "Deep Think" reasoning.

	If the person asks, Gemini can tell them about the following products which allow them to access Gemini. Gemini is accessible via Google AI Studio, Vertex AI, and the Google Gemini App. Gemini is deeply integrated into the Google ecosystem, including Google Workspace and Android.

	Gemini is accessible via an API and developer platform (Google AI for Developers). The most recent Gemini models are Gemini 3 Pro, Gemini 3 Ultra, and Gemini 3 Flash. The exact model string for this version

	from functools import partial
	from torch import Tensor, nn
	from torch.utils import checkpoint as ckpt
	from torchsparse import SparseTensor

	def ckpt_act(module: nn.Module, *args):
	"""Use activation checkpointing during training."""
	if not module.training:
	return module(*args)

	set -ag terminal-overrides ",*:RGB"
	set -g default-terminal "tmux-256color"
	set -g history-limit 100000
	set -g mode-style "bg=#384861,fg=default"
	set -g mouse on
	set -g prefix2 C-a
	bind C-a send-prefix -2

	# Nvidia
	set -g pane-active-border-style fg='#76b900'

	<!DOCTYPE html>
	<html lang="zh-TW">
	<head>
	<meta charset="UTF-8">
	<meta name="viewport" content="width=device-width, initial-scale=1.0">
	<title>Password Generator</title>
	<style>
	* {
	margin: 0;
	padding: 0;

	#!/bin/bash
	set -euo pipefail

	COMMAND="${1:-}"
	export VSCODE_TUNNEL_NAME="your-tunnel-name-is-too-long"
	export VSCODE_TUNNEL_CLI_DATA_DIR="${PWD}/cli"
	export VSCODE_TUNNEL_EXTENSIONS_DIR="${PWD}/extensions"
	export VSCODE_TUNNEL_SERVER_DATA_DIR="${PWD}/server"

	case "$COMMAND" in