scrolls-setup

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Setting up docs/.scrolls/

搭建docs/.scrolls/

docs/.scrolls/

is a small set of markdown files that act as a project's working memory across sessions: what it does, what state it's in, what's known-missing and why, what's next, and what traps to avoid. A

CLAUDE.md

pointer sends every future session to

docs/.scrolls/STARTER.md

first, so state gets picked up instead of re-discovered from scratch each time. This skill scaffolds that system for a project that doesn't have it yet.

This skill's own

assets/templates/

directory holds the source templates — copy from there, never edit those files in place.

docs/.scrolls/

是一组Markdown文件，作为项目跨会话的工作内存：记录项目功能、当前状态、已知缺失内容及原因、下一步计划，以及需要规避的陷阱。

CLAUDE.md

指针会引导所有后续会话先读取

docs/.scrolls/STARTER.md

，这样就能承接已有状态，而非每次从头重新梳理。本工具可为尚未搭建该系统的项目生成这套框架。

本工具的

assets/templates/

目录存放着源模板文件——请从该目录复制模板，切勿直接编辑其中的文件。

Cross-platform

跨平台支持

Unlike the other four scrolls skills, this one has no bundled

.sh

.ps1

script to choose between — nothing here needed porting. File creation happens through your own Read/Write/Edit tools (not raw shell commands), and the one external command this skill relies on (

git rev-parse --show-toplevel

) behaves identically whether invoked from bash or PowerShell — both support the same

$(...)

command-substitution syntax used throughout this file. Works the same on macOS, Linux, and Windows without any environment-specific branching.

与其他四个scrolls工具不同，本工具无需捆绑

.sh

.ps1

脚本进行选择——这里没有需要移植的内容。文件创建通过你自身的读写编辑工具完成（而非原生Shell命令），本工具依赖的唯一外部命令

git rev-parse --show-toplevel

在bash和PowerShell中的表现完全一致——两者均支持本文档中使用的

$(...)

命令替换语法。在macOS、Linux和Windows环境下的运行效果完全相同，无需针对不同环境做分支处理。

Options

选项参数

The user may pass these after

/scrolls-setup

as plain text, in any order — there's no real argv parser here, so read the invocation text yourself and pull out:

-p <path>
/
--path=<path>
/
--path <path>
— a custom docs folder, relative to the current working directory unless given as an absolute path (starts with
```
/
```
). Use this for monorepos or non-standard layouts, e.g.
```
--path=packages/api/docs
```
puts the scrolls at
```
packages/api/docs/.scrolls
```
. This is the one option where you're naming the docs folder directly rather than picking a base directory — see the CLAUDE.md placement note in step 4 for the tradeoff that comes with going deep.
-t
/
--reporoot
— pin everything to the git repository's top level (
```
$(git rev-parse --show-toplevel)
```
), regardless of which subdirectory you actually invoked this from. Fails with a clear message if the current directory isn't inside a git repository — there's no repo root to find.
-l
/
--local
— pin everything to the current working directory explicitly. This is what happens by default anyway when none of
```
-p
```
/
```
-t
```
/
```
-l
```
are given — the flag exists so you (or the user) can say so on purpose, e.g. to skip the mismatch check in step 1.
-r
/
--recurse
— before creating anything, scan recursively under
```
BASE_DIR
```
for a scrolls folder that already exists somewhere nearby (same bounded, pruned,
```
STARTER.md
```
-guarded search
```
/scrolls-hide
```
/
```
/scrolls-unhide
```
use), so a duplicate isn't created by accident. Doesn't change where the new scrolls folder goes if you proceed — see step 1.
-u
/
--unhide
— name the scrolls folder
```
scrolls
```
instead of the default
```
.scrolls
```
, so it isn't dotfile-hidden. Omit for the default (hidden).

-p

-t

, and

-l

are three different ways to answer the same question ("where does the

docs

folder go?") — pass at most one. If more than one is given, stop and ask which was meant.

-r

is independent and combines freely with any of them (or with none).

用户可在

/scrolls-setup

后以纯文本形式传递以下参数，顺序不限——此处没有真正的argv解析器，需自行读取调用文本并提取参数：

-p <path>
/
--path=<path>
/
--path <path>
—— 指定自定义docs文件夹路径，若未给出绝对路径（以
```
/
```
开头）则相对于当前工作目录。适用于单体仓库或非标准目录结构，例如
```
--path=packages/api/docs
```
会将scrolls系统放置在
```
packages/api/docs/.scrolls
```
。这是唯一直接指定docs文件夹的选项，而非选择基础目录——请查看步骤4中关于CLAUDE.md放置位置的说明，了解深层路径带来的权衡。
-t
/
--reporoot
—— 将所有内容固定到Git仓库顶层目录（
```
$(git rev-parse --show-toplevel)
```
），无论实际调用时处于哪个子目录。如果当前目录不在Git仓库内，会给出明确错误提示——此处没有可固定的仓库根目录；建议使用
```
--path
```
替代。
-l
/
--local
—— 明确将所有内容固定到当前工作目录。当未指定
```
-p
```
/
```
-t
```
/
```
-l
```
时，默认就是此行为——该标志的存在是为了让你（或用户）可以明确指定，例如跳过步骤1中的路径不匹配检查。
-r
/
--recurse
—— 在创建任何内容前，递归扫描
```
BASE_DIR
```
下是否已存在scrolls文件夹（使用与
```
/scrolls-hide
```
/
```
/scrolls-unhide
```
相同的有限制、可修剪、由STARTER.md保护的搜索逻辑），避免意外创建重复系统。即使找到现有文件夹，也不会改变新scrolls文件夹的创建位置——请查看步骤1。
-u
/
--unhide
—— 将scrolls文件夹命名为
```
scrolls
```
而非默认的
```
.scrolls
```
，使其不再作为隐藏文件。不指定该参数则使用默认的隐藏文件夹。

-p

、

-t

和

-l

是回答同一问题（“docs文件夹放在哪里？”）的三种不同方式——最多指定其中一个。如果指定了多个，需停止操作并询问用户具体意图。

-r

是独立参数，可与任意一个（或不与任何）参数组合使用。

Steps

操作步骤

1. Resolve BASE_DIR and don't clobber existing work

1. 解析BASE_DIR，避免覆盖现有工作内容

Compute

BASE_DIR

— the directory that will contain both the

docs

folder and

CLAUDE.md

-t
/
--reporoot
given:
```
BASE_DIR = $(git rev-parse --show-toplevel)
```
. If that command fails (not inside a git repository), stop and tell the user — there's no repo root to pin to here; suggest
```
--path
```
instead.
-l
/
--local
given, or -p
/
--path
given: skip straight to computing
```
DOCS_BASE
```
below —
```
-l
```
uses
```
BASE_DIR = $(pwd)
```
and
```
-p
```
bypasses
```
BASE_DIR
```
entirely (see the note under
```
DOCS_BASE
```
).
Nothing given (the common case):
```
BASE_DIR = $(pwd)
```
. But first, if the current directory is inside a git repository, run
```
git rev-parse --show-toplevel
```
and compare it to
```
$(pwd)
```
. If they're the same, or this isn't a git repo, there's nothing to flag. If they differ, tell the user plainly: running from here will create the scrolls at
```
$(pwd)/docs/.scrolls
```
, separate from anything that might already exist at the repository root (
```
<repo-root>
```
), and ask whether that's what they want — the current directory (the default; proceed with it if there's no strong preference either way) or the repo root instead (equivalent to re-running with
```
-t
```
). Don't block indefinitely on this — cwd wins if it's a toss-up, since that's this skill's documented default.

-r

--recurse

was given, do this scan next, before touching the filesystem: search recursively under

BASE_DIR

(skip this if

-p

was given — a custom

--path

is an explicit, deliberate location, not something to second-guess) for any directory named

.scrolls

scrolls

containing a

STARTER.md

, pruning the same heavy/vendor directories

/scrolls-hide

/scrolls-unhide

prune (

node_modules

.git

vendor

dist

build

.venv

venv

__pycache__

target

.next

.cache

), bounded to a reasonable depth (8 is what the other scrolls skills use). If this finds an existing scrolls folder anywhere under

BASE_DIR

other than the exact

SCROLLS_PATH

you're about to create, surface it and ask whether the user meant to run

/scrolls-update

against that one instead of creating a new, separate one here — proceed with creation only if they confirm that's what they want (e.g. a deliberately separate scrolls system for a sub-project).

Then:

DOCS_BASE    = the --path value if -p was given, else "${BASE_DIR}/docs"
SCROLLS_DIR  = "scrolls" if --unhide/-u was passed, else ".scrolls"
SCROLLS_PATH = "${DOCS_BASE}/${SCROLLS_DIR}"

SCROLLS_PATH

replaces every

docs/.scrolls

you'll see referenced below and in the templates — the rest of this skill talks about "the scrolls path" generically. If the user already has a hidden

.scrolls

set up and wants it converted to visible later, that's a separate, dedicated operation — point them at

/scrolls-unhide

rather than re-running this skill.

SCROLLS_PATH

already exists with files in it, stop and ask the user whether they want you to fill in only the missing files or leave it alone — never overwrite an existing scroll file silently, since

HANDOFF.md

PLAN.md

/etc. may hold real accumulated state. The same caution applies to

CLAUDE.md

: never blow away existing content. If

--unhide

was passed but a

.scrolls

folder already exists there (or vice versa), don't create a second, parallel scrolls folder — tell the user and point at

/scrolls-unhide

instead.

计算

BASE_DIR

——即包含

docs

文件夹和

CLAUDE.md

的目录：

指定了
-t
/
--reporoot
：
```
BASE_DIR = $(git rev-parse --show-toplevel)
```
。如果该命令执行失败（当前目录不在Git仓库内），停止操作并告知用户——此处没有可固定的仓库根目录；建议使用
```
--path
```
替代。
指定了
-l
/
--local
，或指定了
-p
/
--path
：直接跳至下文计算
```
DOCS_BASE
```
——
```
-l
```
使用
```
BASE_DIR = $(pwd)
```
，
```
-p
```
则完全绕过
```
BASE_DIR
```
（请查看
```
DOCS_BASE
```
下的说明）。
未指定任何参数（常见情况）：
```
BASE_DIR = $(pwd)
```
。但首先，如果当前目录在Git仓库内，执行
```
git rev-parse --show-toplevel
```
并与
```
$(pwd)
```
对比。如果两者相同，或当前目录不在Git仓库内，则无需提示。如果两者不同，需明确告知用户：在此处运行会在
```
$(pwd)/docs/.scrolls
```
创建scrolls系统，与仓库根目录（
```
<repo-root>
```
）下可能已存在的系统相互独立，并询问用户是否确定如此——当前目录为默认选项；若用户无明确偏好，则使用当前目录继续操作，这是本工具文档中规定的默认行为。

如果指定了

-r

--recurse

，则在操作文件系统前执行以下扫描：递归搜索

BASE_DIR

下是否存在名为

.scrolls

或

scrolls

且包含

STARTER.md

的目录（跳过与

/scrolls-hide

/scrolls-unhide

相同的大型/供应商目录：

node_modules

、

.git

、

vendor

、

dist

、

build

、

.venv

、

venv

、

__pycache__

、

target

、

.next

、

.cache

），搜索深度限制在合理范围内（其他scrolls工具使用的深度为8）。如果在

BASE_DIR

下除了即将创建的

SCROLLS_PATH

之外的位置找到现有scrolls文件夹，需告知用户并询问是否要针对该现有系统运行

/scrolls-update

，而非在此处创建新的独立系统——仅当用户确认需要创建新系统时（例如为子项目创建独立的scrolls系统），才继续执行创建操作。

然后：

DOCS_BASE    = 若指定了-p则为其路径值，否则为"${BASE_DIR}/docs"
SCROLLS_DIR  = 若指定了--unhide/-u则为"scrolls"，否则为".scrolls"
SCROLLS_PATH = "${DOCS_BASE}/${SCROLLS_DIR}"

SCROLLS_PATH

替代下文及模板中所有提及的

docs/.scrolls

——本工具后续将泛称其为“scrolls路径”。如果用户已设置隐藏的

.scrolls

文件夹，之后想要转换为可见文件夹，这是一个单独的专用操作——请引导用户使用

/scrolls-unhide

，而非重新运行本工具。

如果

SCROLLS_PATH

已存在且包含文件，需停止操作并询问用户是仅填充缺失文件还是保留现有内容——切勿静默覆盖现有scroll文件，因为

HANDOFF.md

PLAN.md

等文件可能包含真实的累积状态。对

CLAUDE.md

也需同样谨慎：切勿删除现有内容。如果指定了

--unhide

但该位置已存在

.scrolls

文件夹（反之亦然），请勿创建第二个并行的scrolls文件夹——告知用户并引导其使用

/scrolls-unhide

。

2. Gather just enough project context

2. 收集必要的项目上下文

Don't interview the user at length — a minimal setup should be fast. Infer what you can in a few seconds:

Project name: from
```
package.json
```
's
```
name
```
,
```
pyproject.toml
```
's
```
[project].name
```
,
```
Cargo.toml
```
, or failing that the directory's basename.
One-line tagline: from the same manifest's
```
description
```
field, or the first line of an existing
```
README.md
```
, or omit it — it's optional flavor text, not load-bearing.
Quick orientation: one short paragraph on the project's actual shape — main entry point(s), the one or two files/directories that matter most, and the primary language/stack. Get this by a quick look at the repo root and manifest files, not a deep audit. If the project is genuinely empty (brand new, no code yet), say so plainly rather than inventing structure.

If any of this is ambiguous (e.g. a monorepo with several

package.json

s), a single clarifying question is fine — but default to acting rather than blocking on questions the repo already answers.

无需对用户进行冗长访谈——最小化搭建应快速完成。在几秒内尽可能推断以下信息：

项目名称：来自
```
package.json
```
的
```
name
```
字段、
```
pyproject.toml
```
的
```
[project].name
```
字段、
```
Cargo.toml
```
，若均无法获取则使用目录名称。
一行标语：来自同一配置文件的
```
description
```
字段，或现有
```
README.md
```
的第一行，若无法获取则省略——这是可选的修饰文本，非必需内容。
快速介绍：关于项目实际结构的简短段落——主入口文件、最重要的一两个文件/目录，以及主要语言/技术栈。通过快速查看仓库根目录和配置文件获取，无需深度审计。如果项目完全为空（全新项目，无代码），则直接说明，无需虚构结构。

如果某些信息不明确（例如单体仓库包含多个

package.json

），可提出一个澄清问题——但优先执行操作，而非因仓库已能回答的问题而停滞。

3. Create the seven files

3. 创建七个文件

Copy each file from

assets/templates/

into

SCROLLS_PATH

on disk, substituting the

{{PROJECT_NAME}}

{{PROJECT_TAGLINE}}

{{QUICK_ORIENTATION}}

, and

{{SCROLLS_PATH}}

placeholders with what you gathered in step 2 and computed above.

SCROLLS_PATH

is an absolute filesystem path when it came from

-t

-l

/the default (all resolve through

BASE_DIR

, which is always absolute) — that's fine for the actual file writes, but the text you substitute for
{{SCROLLS_PATH}}
inside the templates is not the same string:

-t
/
-l
/default (
BASE_DIR
-derived): substitute the short form,
```
docs/${SCROLLS_DIR}
```
(e.g.
```
docs/.scrolls
```
) — never the absolute
```
SCROLLS_PATH
```
. This is what keeps the files portable:
```
CLAUDE.md
```
always ends up living at
```
BASE_DIR
```
too (step 4), so a reference relative to
```
BASE_DIR
```
is correct regardless of whether
```
BASE_DIR
```
was
```
$(pwd)
```
or the git root, and regardless of which machine or clone reads it later. Baking in an absolute path here would break the moment the repo is cloned somewhere else.
-p
/
--path
: substitute the full
```
DOCS_BASE
```
-based
```
SCROLLS_PATH
```
as given, unchanged from before (e.g.
```
packages/api/docs/.scrolls
```
) — this path is already relative, and where
```
CLAUDE.md
```
ends up for this case is judgment-dependent (see step 4), so keep the existing behavior rather than guessing at a shorter form.

Template	→	Purpose
`STARTER.md`	`SCROLLS_PATH/STARTER.md`	Reading order + when-to-update table. The entry point every session reads first.
`SPEC.md`	`SCROLLS_PATH/SPEC.md`	Feature list, filled in as features ship.
`HANDOFF.md`	`SCROLLS_PATH/HANDOFF.md`	Snapshot of current state — overwritten each session, not appended.
`GAP_ANALYSIS.md`	`SCROLLS_PATH/GAP_ANALYSIS.md`	What's known-missing or partial.
`GAP_CONTEXT.md`	`SCROLLS_PATH/GAP_CONTEXT.md`	Why each gap exists (deliberate cut vs. oversight vs. blocker).
`PLAN.md`	`SCROLLS_PATH/PLAN.md`	Prioritized, ticketed backlog.
`WISDOM.md`	`SCROLLS_PATH/WISDOM.md`	Constraints / Traps / Ditches / Wisdom sections.

This is the minimal set — exactly the six files

STARTER.md

walks through, plus

STARTER.md

itself. Don't invent extra scroll files (security reviews, architecture-decision records, subsystem deep-dives) up front; those get added later, organically, by whoever's doing that specific work, following the pattern

STARTER.md

's own last section describes. Leave

{{PROJECT_TAGLINE}}

blank (drop the placeholder entirely, don't leave literal

{{...}}

text) if you found nothing worth using — it reads fine as

You're picking up work on **Foo**.

with no tagline clause.

STARTER.md

and

CLAUDE_MD_BLOCK.md

are the only templates containing

{{SCROLLS_PATH}}

— substitute it per the rule above (short form for

-t

-l

/default, full

SCROLLS_PATH

for

-p

), not a placeholder string, and use the same substitution in both files.

Templates are intentionally close to empty (placeholder bullets like "(none tracked yet)") — resist the urge to pre-populate

SPEC.md

with a guessed feature list or

PLAN.md

with invented tickets. A minimal scaffold's job is to hold the shape; the content accumulates from real sessions. The one exception is

STARTER.md

's "Quick orientation" section, which is worth getting right since it's the one piece of static context every session leans on immediately.

将

assets/templates/

中的每个文件复制到磁盘上的

SCROLLS_PATH

，并将

{{PROJECT_NAME}}

、

{{PROJECT_TAGLINE}}

、

{{QUICK_ORIENTATION}}

和

{{SCROLLS_PATH}}

占位符替换为步骤2收集的信息和上文计算的路径。当

SCROLLS_PATH

来自

-t

-l

/默认设置（均通过

BASE_DIR

解析，始终为绝对路径）时，文件写入使用绝对路径没问题，但模板中替换
{{SCROLLS_PATH}}
的文本并非同一字符串：

-t
/
-l
/默认设置（基于BASE_DIR）：替换为短格式
```
docs/${SCROLLS_DIR}
```
（例如
```
docs/.scrolls
```
）——切勿使用绝对路径
```
SCROLLS_PATH
```
。这样可保持文件的可移植性：
```
CLAUDE.md
```
也会存放在
```
BASE_DIR
```
（步骤4），因此相对于
```
BASE_DIR
```
的引用始终正确，无论
```
BASE_DIR
```
是
```
$(pwd)
```
还是Git根目录，也无论后续在哪个机器或克隆版本中读取。在此处写入绝对路径会导致仓库克隆到其他位置时失效。
-p
/
--path
：直接使用之前计算的基于
```
DOCS_BASE
```
的完整
```
SCROLLS_PATH
```
（例如
```
packages/api/docs/.scrolls
```
）——该路径已为相对路径，且此情况下
```
CLAUDE.md
```
的存放位置需根据实际情况判断（请查看步骤4），因此保持现有行为，而非猜测更短的路径格式。

模板文件	目标文件	用途
`STARTER.md`	`SCROLLS_PATH/STARTER.md`	阅读顺序及更新时机表。每个会话首先读取的入口文件。
`SPEC.md`	`SCROLLS_PATH/SPEC.md`	功能列表，随功能交付逐步填充。
`HANDOFF.md`	`SCROLLS_PATH/HANDOFF.md`	当前状态快照——每次会话覆盖更新，而非追加。
`GAP_ANALYSIS.md`	`SCROLLS_PATH/GAP_ANALYSIS.md`	已知缺失或未完成的内容。
`GAP_CONTEXT.md`	`SCROLLS_PATH/GAP_CONTEXT.md`	每个缺失内容存在的原因（故意删减、疏忽或阻塞）。
`PLAN.md`	`SCROLLS_PATH/PLAN.md`	已排序、带工单的待办事项列表。
`WISDOM.md`	`SCROLLS_PATH/WISDOM.md`	约束条件/陷阱/误区/经验总结部分。

这是最小化的文件集合——恰好是

STARTER.md

中提及的六个文件，外加

STARTER.md

本身。请勿预先创建额外的scroll文件（安全审查、架构决策记录、子系统深度分析等）；这些文件后续可由相关人员根据

STARTER.md

最后一节描述的模式有机添加。如果未找到合适的

{{PROJECT_TAGLINE}}

，则留空（直接删除占位符，不要保留字面的

{{...}}

文本）——即使没有标语，

You're picking up work on **Foo**.

这样的表述也完全通顺。

只有

STARTER.md

和

CLAUDE_MD_BLOCK.md

模板包含

{{SCROLLS_PATH}}

——需按照上述规则替换（

-t

-l

/默认设置使用短格式，

-p

使用完整

SCROLLS_PATH

），而非保留占位符字符串，且两个文件需使用相同的替换内容。

模板文件故意设计为接近空白（例如占位符项目符号"(none tracked yet)"）——请勿预先填充

SPEC.md

的猜测功能列表或

PLAN.md

的虚构工单。最小化框架的作用是提供结构；内容需从真实会话中逐步积累。唯一例外是

STARTER.md

的“快速介绍”部分，值得准确填写，因为这是每个会话立即依赖的静态上下文。

4. Point CLAUDE.md at STARTER.md

4. 配置CLAUDE.md指向STARTER.md

Read

assets/templates/CLAUDE_MD_BLOCK.md

and substitute

{{SCROLLS_PATH}}

in it the same way as

STARTER.md

— that's the block to install.

-t
/
-l
/default (
BASE_DIR
-derived):
```
CLAUDE.md
```
goes at
```
BASE_DIR
```
— the same directory that now contains
```
docs
```
. This is fixed and unambiguous:
```
BASE_DIR
```
is exactly what
```
-t
```
/
```
-l
```
/the mismatch check in step 1 resolved, so there's no separate "project root" judgment call to make here anymore.
-p
/
--path
: where
```
CLAUDE.md
```
belongs is genuinely judgment-dependent, since a custom
```
--path
```
might point at an independent monorepo package (its own
```
CLAUDE.md
```
, short local references — closer to what
```
-l
```
would produce if you'd
```
cd
```
ed into that package first) or be one piece of a larger repo meant to stay under a single root
```
CLAUDE.md
```
referencing the full path. Use whatever
```
CLAUDE.md
```
location the project's other tooling already expects; if genuinely unclear, ask rather than guessing — this is the one case where a wrong guess is expensive (a broken reference baked into checked-in docs).

Once you know where

CLAUDE.md

goes:

No
CLAUDE.md
there yet: create one containing exactly that block.
CLAUDE.md
exists but has no scrolls-path reference: insert the block near the top of the file (before other instructions, since "read this first" only works if it's read first), separated by blank lines from surrounding content. If the file already opens with its own top-level heading, add the block's heading as a subsection instead of a second top-level
```
# Project instructions
```
— match the existing file's heading structure rather than fighting it.
CLAUDE.md
already references
SCROLLS_PATH/STARTER.md
(or the short form, if that's what applies here): leave it alone; note this to the user instead of duplicating the block.

读取

assets/templates/CLAUDE_MD_BLOCK.md

并按照与

STARTER.md

相同的规则替换

{{SCROLLS_PATH}}

——这就是需要添加的内容块。

-t
/
-l
/默认设置（基于BASE_DIR）：
```
CLAUDE.md
```
存放在
```
BASE_DIR
```
——即现在包含
```
docs
```
的目录。这是固定且明确的：
```
BASE_DIR
```
已由
```
-t
```
/
```
-l
```
/步骤1中的路径不匹配检查解析完成，因此无需再判断“项目根目录”。
-p
/
--path
：
```
CLAUDE.md
```
的存放位置需根据实际情况判断，因为自定义
```
--path
```
可能指向独立的单体仓库包（需单独的
```
CLAUDE.md
```
和短本地引用——类似于进入该包后使用
```
-l
```
的效果），或作为大型仓库的一部分，需保留在单一根目录
```
CLAUDE.md
```
中并引用完整路径。使用项目其他工具已采用的
```
CLAUDE.md
```
存放位置；若确实不明确，需询问用户而非猜测——此情况下错误猜测的代价较高（会将无效引用写入已提交的文档）。

确定

CLAUDE.md

的存放位置后：

该位置尚无CLAUDE.md：创建文件并写入上述内容块。
CLAUDE.md已存在但未引用scrolls路径：将内容块插入文件顶部附近（在其他说明之前，因为“优先阅读”只有在最前面才有效），与周围内容用空行分隔。如果文件已以一级标题开头，则将内容块的标题作为子标题，而非添加第二个一级标题
```
# Project instructions
```
——匹配现有文件的标题结构，而非强行修改。
CLAUDE.md已引用
SCROLLS_PATH/STARTER.md
（或适用的短格式）：保留现有内容；告知用户无需重复添加内容块。

5. Report back

5. 反馈结果

Summarize what was created vs. what already existed and was left untouched, and name the two or three things most worth the user's attention next: filling in

STARTER.md

's quick-orientation paragraph if you had to guess at it, and writing the first real

SPEC.md

entry once something ships.

总结已创建的内容与已存在且未修改的内容，并指出用户接下来最值得关注的两三个事项：如果之前猜测了

STARTER.md

的快速介绍段落，需完善该内容；当有功能交付时，写入第一条真实的

SPEC.md

条目。