1.3k kharmanskyi

os-step-by-step Skill

>- 当需要用户采取行动——运行命令、粘贴密钥、点击、批准——以及用户问某事怎么做或表示不知道该做什么时:"一步一步来""带我过一遍""我该干嘛""我该怎么办""我不明白要做什么""解释一下我需要做什么",任何语言都必须调用此技能。挑选下一项任务请用 os-whats-next;此技能用于完成眼前的事。先让请求站得住脚:自己先试试、找别的路子、把请求缩小到只有用户能做的那部分。然后每步一个动作,命令标注其影响范围。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

os-step-by-step

The user is not stuck because the task is hard - they cannot tell what they
are being asked to do. This skill turns "I need something from you" into an
instruction a person who does not read code can follow without a follow-up
question. os-done-or-not reports that something is needed; this one says
exactly how.

Language

Write in the language the user speaks in this session, detected from the
conversation. Commands, file names and identifiers stay English.

When to use

Triggers live in the description above - any moment you need the user's hands,
or they ask what to do.

Step 0 - earn the right to ask

Every ask costs the user a context switch. Prove it is necessary; stop at the
first item that clears the block:

  1. Try it. A real 403 is a finding; "I probably lack permission" is a guess.
  2. Find another route. Another tool, a value already on the host, a file

you can read.

  1. Shrink the ask. Obtain what you can yourself; hand over only the

irreducible part.

  1. Check you are not asking twice - search the session and the reports

folder first.

Three walls where asking IS the correct move, never to be worked around:
pulling a secret into your own context, loosening a guard that is there on
purpose, doing what the user said only they do.

The shape

Always this order - the ask first, never after the diagnosis.

**What I need from you: <one sentence, plain words>.**

Why you and not me: <one or two sentences. A real reason, in human terms.>

**Step 1. <action in three to six words>**
<What to do. One action only.>

**Step 2. <action>**
<...>

**How you'll know it worked.**
<What the user will see. What to do if they see something else.>

**What happens next, on my side.**
<One line: what you do once they are done, and what you will say.>

After they act - verify, do not trust

"Done" is a claim. Run the one quickest check that would fail if the step had
not worked - the file exists with the right owner, the service answers, the
value works once. Never print a secret to confirm it: confirm its effect. If
the check fails, give only the corrected step - never the whole list again.
Verified versus assumed is exactly what os-done-or-not needs for "yes"
versus "not checked".

Secrets and dangerous steps

A secret never goes through the chat - it would stay in the history and the
logs. The command below puts it where it belongs directly; say who deletes it
and when. A hard-to-undo step - live users, money, deletion - gets its own
warning line before the command.

Writing a secret to a remote host - one line: it prompts, hides the
typing, refuses a truncated paste, confirms by size:

printf 'Paste the connection string, then press Enter: '; IFS= read -rs V; echo; if [ ${#V} -lt 20 ]; then echo "Only ${#V} characters - that looks truncated, nothing was saved."; else printf '%s' "$V" | ssh root@HOST 'umask 077 && cat > /path/to/secret' && ssh root@HOST 'echo "Saved, $(wc -c < /path/to/secret) bytes"'; fi; unset V

Typed by hand rather than pasted - ask twice; a typo in a hidden field is
otherwise undetectable:

printf 'Enter the token: '; IFS= read -rs A; echo; printf 'Enter it again: '; IFS= read -rs B; echo; if [ "$A" != "$B" ]; then echo "The two entries differ - nothing saved, run it again."; else printf '%s' "$A" | (umask 077; cat > /path/to/secret) && echo "Saved, $(wc -c < /path/to/secret) bytes"; fi; unset A B

These templates are for bash and zsh and have not been run on Windows. There,
ask the user to open a separate Git Bash window just for this command, and to
leave the agent running where it is. Never write an unchecked PowerShell one.

Two properties the templates cannot keep for you, both measured:

  • printf …; IFS= read -rs VAR - never read -rsp. In zsh -p means

"read from a coprocess": the variable comes back empty with no error and the
secret file is written blank. macOS defaults to zsh.

  • The value never appears in the command itself - only piped from the

variable; anything in the arguments lands in shell history and the process
list.

The rest the templates already embody - one single line, umask 077 before
writing, refuse short input, confirm by byte count never by content, unset
at the end. Adapt the prompt, the threshold and the path; keep every property.

The same pattern serves any value the user must supply by hand - a public
key, a domain, an address, an id. Prompt for it the same way; keep the typing
visible when the value is not secret (drop -s), skip the length gate when
short is valid - and always end with a plain-words confirmation of what just
happened, so pressing Enter never feels like dropping a coin into a well.

Choices, not instructions

A decision gets no steps. Use your tool's question picker where it has one
(in Claude Code, AskUserQuestion), with the pack's contract: plain question,
why it matters, what changes later, easy to undo, two to four options, the
recommended one first and marked. Where there is no picker, write the same
question and options as plain text, the recommended one first and marked.

Hard rules

  1. The ask goes first. What you tried and what failed is your problem -

one line at the end, or nothing.

  1. One action per step. Two commands is two steps.
  2. Label every command with what it touches - the test server, the live

server, their own machine. Look-alike steps on different targets: say what
happens if they are swapped.

  1. A command is self-contained. One line the user pastes and runs; if it

needs a value, it asks for it. Never make the user feed input by
redirection, a heredoc or Ctrl-D.

  1. No jargon inside a step. Avoid terms instead of explaining them: write

what the person sees and clicks. One unavoidable term may stay - without a
lecture.

  1. Always give a way to check. A step the user cannot verify is a step

they will redo out of doubt.

  1. Never mix your work with theirs. Two lists with plain headings; a

buried "this one's on you" is not an instruction.

  1. Never compress a multi-step sequence - here brevity causes misreads.

Blocks of two to three sentences; longer gets skimmed, and a skimmed step
is a missed step.

Known gotchas

  • A heading is not a summary: "one command per host" above two commands reads

as one command. Count out loud.

  • A dropped step is "skip step 3" - never a silent renumber.
  • If you can verify it yourself, verify it yourself; do not ask for

confirmation you do not need.

  • "Say done" needs a subject - the user may have three of your requests open.