arc-skill

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Arc skill

Arc技能

Set the launcher once, work inside one directory per game, then start or resume:

bash

ARC="<this-skill-directory>/scripts/arc"   # absolute path
mkdir -p <run-dir> && cd <run-dir>         # one directory = one run
"$ARC" start <GAME_ID>

start

defaults to a local simulator with competition semantics: one logical run, append-only history, paid resets, completed levels never lost. It is crash-safe — after any interruption, rerun

"$ARC" start <GAME_ID>

and the run is resumed or replayed exactly. Use

--mode competition

(live remote server, ~15-minute idle lease, no replay recovery) only when the user explicitly asks. Never create a second run for the same game, never inspect the game's source or private state, and never edit

.arc/

by hand except

NOTES.md

只需配置一次启动器，每个游戏使用单独目录，即可启动或恢复游戏：

bash

ARC="<this-skill-directory>/scripts/arc"   # 绝对路径
mkdir -p <run-dir> && cd <run-dir>         # 一个目录对应一次运行
"$ARC" start <GAME_ID>

start

默认使用具有竞赛语义的本地模拟器：一次逻辑运行、仅追加历史记录、重置需消耗点数、已完成关卡不会丢失。它具备崩溃安全性——无论任何中断，重新运行

"$ARC" start <GAME_ID>

即可精确恢复或重放运行。仅当用户明确要求时，才使用

--mode competition

模式（远程实时服务器、约15分钟空闲租期、不支持重放恢复）。切勿为同一游戏创建第二次运行，切勿查看游戏源码或私有状态，除

NOTES.md

外，切勿手动编辑

.arc/

目录。

The game

游戏介绍

A 64×64 board of 16 colors, hidden rules, several levels. Interface priors (not guarantees): ACTION1/2/3/4 = up/down/left/right, ACTION5 = interact, ACTION6:x,y = click at x=column y=row, ACTION7 = undo. Availability can change after every action.

Finish first. Historically every lost point came from unfinished games, not from extra actions. Early levels are usually cheap tutorials: a wrong action that teaches a mechanic beats a minute of deliberation. Act early, act often, learn from every grade. Efficiency only matters once finishing is likely.

游戏是一个64×64的棋盘，包含16种颜色、隐藏规则和多个关卡。界面预设（非绝对）：ACTION1/2/3/4 = 上/下/左/右，ACTION5 = 交互，ACTION6:x,y = 点击x列y行，ACTION7 = 撤销。每次行动后，可用操作可能会变化。

优先完成游戏。 过往所有失分均来自未完成的游戏，而非额外行动。早期关卡通常是低成本教程：一个能理解机制的错误行动，胜过一分钟的深思熟虑。尽早行动、频繁行动，从每次评分中学习。只有在大概率能完成游戏时，效率才重要。

The loop — look, predict, act, compare, note

操作循环——观察、预测、行动、对比、记录

Look: open the printed
```
IMAGE
```
; read the worded
```
TRANSITION
```
story.
Predict + act: every action requires
```
--predict
```
; the harness grades it:

bash

"$ARC" act ACTION1 --predict "move 12,5 0,-1"
"$ARC" act ACTION6 3 14 --predict "cell 3,14=9; region 0:8,10:20" --because "test button"

Compare: read the ✓/✗ grade, the worded
```
TRANSITION
```
story, and the cell-exact
```
DIFF
```
before/after masks in the result (
```
arc view
```
re-renders them for any event). A ✗ is the most valuable thing that can happen — reality just corrected you for one action.
Note: keep
```
.arc/NOTES.md
```
to one page with three sections —
```
Verified (cite event ids)
```
,
```
Assumed / open questions
```
,
```
Plan
```
. After a ✗, fix the notes before the next action.
```
arc status
```
prints the file in full, so it is also your recovery story after any context loss.

Claim vocabulary (full reference:

"$ARC" act --help

noop

change

cell X,Y=V

move X,Y DX,DY

vanish X,Y

region X0:X1,Y0:Y1

level+1

win

; several separated by

. Coordinates are x=column, y=row, like ACTION6. Free text is allowed and merely claims "something changes" — prefer one specific claim; it grades sharper and teaches more.

观察：打开输出的
```
IMAGE
```
；阅读文字版的
```
TRANSITION
```
事件描述。
预测+行动：每次行动都需要添加
```
--predict
```
参数；工具框架会对行动进行评分：

bash

"$ARC" act ACTION1 --predict "move 12,5 0,-1"
"$ARC" act ACTION6 3 14 --predict "cell 3,14=9; region 0:8,10:20" --because "test button"

对比：查看✓/✗评分、文字版
```
TRANSITION
```
事件描述，以及结果中精确到单元格的前后
```
DIFF
```
掩码（
```
arc view
```
可重新渲染任意事件的这些内容）。出现✗是最有价值的情况——现实刚纠正了你的一次行动。
记录：将
```
.arc/NOTES.md
```
保持为单页，包含三个部分——
```
已验证（引用事件ID）
```
、
```
假设/未解决问题
```
、
```
计划
```
。出现✗后，在下次行动前更新笔记。
```
arc status
```
会完整打印该文件，因此它也是你在丢失上下文后的恢复依据。

声明词汇表（完整参考：

"$ARC" act --help

）：

noop

、

change

、

cell X,Y=V

、

move X,Y DX,DY

、

vanish X,Y

、

region X0:X1,Y0:Y1

、

level+1

、

win

；多个声明用

分隔。坐标为x=列，y=行，与ACTION6一致。允许使用自由文本，仅表示“某些内容发生变化”——优先使用具体声明，这样评分更精准，也能学到更多。

Batching proven mechanics

批量执行已验证机制

Once a mechanic is verified, stop paying one command per step — batch with a claim on every step; execution halts at the first miss so a wrong theory cannot burn the rest of the queue:

bash

"$ARC" commit \
  --step "ACTION4 :: move 12,5 1,0" \
  --step "ACTION4 :: move 13,5 1,0" \
  --step "ACTION1 :: level+1"

Batch only movement you can predict cell-exactly or level-exactly; never batch exploration.

一旦某个机制被验证，无需再逐步骤执行命令——可批量执行并为每一步添加声明；执行会在首次失败时停止，避免错误理论消耗后续队列中的行动次数：

bash

"$ARC" commit \
  --step "ACTION4 :: move 12,5 1,0" \
  --step "ACTION4 :: move 13,5 1,0" \
  --step "ACTION1 :: level+1"

仅批量执行可精确预测单元格或关卡变化的移动操作；切勿批量执行探索性操作。

Levels, consumables, reset

关卡、消耗品、重置

Completing a level archives your notes to
```
.arc/levels/
```
. The new level may reuse mechanics — treat every earlier
```
Verified
```
claim as
```
Assumed
```
until it survives one test on the new board (status reminds you until the notes change).
An object that vanished and never came back is a consumable. Spend consumables last, after reversible probes; before an irreversible-looking action, prefer ACTION7 (undo) tests when available.
```
"$ARC" reset --because "<why this board is unrecoverable>"
```
rewinds only the current level, for the price of one action. After
```
GAME_OVER
```
the reason may be omitted. Completed levels and history are never lost.

完成一个关卡后，你的笔记会归档到
```
.arc/levels/
```
。新关卡可能会复用之前的机制——在新棋盘上通过一次测试前，将所有之前的
```
已验证
```
声明视为
```
假设
```
（状态信息会提醒你，直到笔记更新）。
消失后不再出现的物体是消耗品。应在完成可逆探测后再使用消耗品；在执行看似不可逆的行动前，若可用，优先使用ACTION7（撤销）进行测试。
```
"$ARC" reset --because "<说明当前棋盘无法恢复的原因>"
```
仅重置当前关卡，代价是消耗一次行动次数。
```
GAME_OVER
```
后可省略原因。已完成的关卡和历史记录永远不会丢失。

When a level resists — the rules tier (optional)

当关卡难以突破时——规则层（可选）

Status nudges you after many actions or repeated misses on one level. Then, and only then, escalate from prose to executable rules: write a plain

rules.py

(grounding, step, actions, goal), verify it against the entire recorded history, and let A* search find the plan. Each plan step carries its own prediction, so live execution still halts on the first surprise.

bash

"$ARC" rules help      # the compact contract
"$ARC" rules init      # template; then edit rules.py
"$ARC" rules replay    # must fit or gap on every recorded transition
"$ARC" rules solve     # writes .arc/plan.json
"$ARC" commit @.arc/plan.json

Model only verified mechanics; mark everything else

Unknown("why")

— replay reports gaps honestly instead of pretending a fit. This tier is never required and never worth it before the game has taught you its mechanics.

当在一个关卡中执行了多次行动或反复失败时，状态信息会提示你。此时（且仅在此时），可从文字描述升级到可执行规则：编写一个简单的

rules.py

文件（包含基础设定、步骤、行动、目标），对照整个记录的历史进行验证，然后让A*搜索算法找到计划。每个计划步骤都带有自己的预测，因此实时执行仍会在首次出现意外时停止。

bash

"$ARC" rules help      # 简洁的规则约定
"$ARC" rules init      # 生成模板；然后编辑rules.py
"$ARC" rules replay    # 必须匹配所有记录的事件，否则会显示差异
"$ARC" rules solve     # 生成.arc/plan.json文件
"$ARC" commit @.arc/plan.json

仅对已验证的机制建模；将所有其他内容标记为

Unknown("原因")

——重放会如实报告差异，而非假装匹配。规则层并非必需，且在游戏尚未让你理解其机制前使用毫无价值。

Evidence tools

证据工具

bash

"$ARC" status                       # full picture + notes; run after context loss
"$ARC" view --grid                  # exact 0-f pixels
"$ARC" view --crop 8:24,10:30       # exact half-open crop
"$ARC" view --event 12 --frames     # animation frames of a past action
"$ARC" python 'connected_components(grid)'
"$ARC" python 'shortest_path((r0,c0),(r1,c1), passable_mask)'

arc python

preloads every settled board (

grid

previous

frames

transitions

actions

), NumPy, perception helpers, BFS, and A* — free offline computation; use it for parsing and pathfinding instead of paid probing.

bash

"$ARC" status                       # 完整状态+笔记；丢失上下文后运行
"$ARC" view --grid                  # 精确显示0-f像素
"$ARC" view --crop 8:24,10:30       # 精确裁剪半开区间区域
"$ARC" view --event 12 --frames     # 显示过往行动的动画帧
"$ARC" python 'connected_components(grid)'
"$ARC" python 'shortest_path((r0,c0),(r1,c1), passable_mask)'

arc python

会预加载所有已稳定的棋盘数据（

grid

、

previous

、

frames

、

transitions

、

actions

）、NumPy、感知辅助工具、BFS和A*算法——这是免费的离线计算；可用于解析和路径查找，而非消耗点数进行探测。

Install location

安装位置

This folder is agent-agnostic. Copy it into the platform's auto-discovered skill directory (Claude Code:

.claude/skills/arc-skill/

, Codex:

.agents/skills/arc-skill/

). From an unrecognized location, instruct the agent to read this

SKILL.md

completely before starting.

此文件夹与Agent无关。将其复制到平台自动识别的技能目录中（Claude Code：

.claude/skills/arc-skill/

，Codex：

.agents/skills/arc-skill/

）。若处于未被识别的位置，需指示Agent在启动前完整阅读此

SKILL.md

文件。