- name
SKILL.md
스킬 파일은 게시자의 저장소에 그대로 있습니다.
복사한 명령을 터미널에 붙여넣습니다. 이 SKILL.md가 해당 에이전트에 설치됩니다.
npx skills add kharmanskyi/open-steps --skill os-what-could-go-wrong -a cursor -y복사한 명령을 터미널에 붙여넣습니다. 이 SKILL.md가 해당 에이전트에 설치됩니다.
npx skills add kharmanskyi/open-steps --skill os-what-could-go-wrong -a claude-code -y복사한 명령을 터미널에 붙여넣습니다. 이 SKILL.md가 해당 에이전트에 설치됩니다.
npx skills add kharmanskyi/open-steps --skill os-what-could-go-wrong -a codex -y복사한 명령을 터미널에 붙여넣습니다. 이 SKILL.md가 해당 에이전트에 설치됩니다.
npx skills add kharmanskyi/open-steps --skill os-what-could-go-wrong -a grok -y복사한 메시지를 Grok Bot 채팅에 보냅니다. Bot이 GitHub에서 이 SKILL.md를 가져와 모든 Bot이 공유하는 스킬에 추가합니다.
Install this skill: https://github.com/kharmanskyi/open-steps/tree/HEAD/skills/os-what-could-go-wrong게시자 제공
Assume it already failed, then work backwards to find out why - while there is still time to change it. The attack runs in a fresh agent that had no part in the decision, because an agent that helped shape one reviews it far too gently: it defends its own reasoning, and it misses the thing that kills the plan out of politeness.
os-ask-simple screens a choice before the user picks one. This one runs
after the choice is made and before it can no longer be taken back.
Write in the language the user speaks in this session, detected from the conversation. Names, figures and identifiers stay as they are. The fresh agent cannot see this conversation, so the language travels in the handover.
Find the decision first. Given as text or a document, that is it. Asked at the end of a discussion, it is the decision the discussion arrived at, and say which one you took it to be. If neither, ask which decision to attack.
Then fill every line. This is the only thing the fresh agent will ever see.
DECISION BRIEF
What will be done: three to seven sentences
What it is for: the problem it solves, and what success looks like, measured
wherever it can be
The main moves: the money, the people, the systems, the dates
What is fixed: the constraints, plus the surrounding facts that matter
Who it lands on: who and what is affected if this goes wrong
What cannot be undone: which parts are one-way
When we would know: the date success or failure actually gets judged
What we know: facts from documents, data, the repository, past incidents,
each with where it came from
What we are assuming: every gap nobody could close, written as an assumption
Close the gaps in this order, and stop as soon as a gap is closed.
~/.claude/open-steps/reports/, what went wrong last time.os-ask-simple, three at most.
Nobody there to answer, or an answer that would not move the verdict -> 3.(assumed), and repeat it under "What we are assuming".The brief states facts and open questions. It never makes the case for the decision: a brief that argues gets a report that agrees.
Pick the depth, say which in one line, and carry on; the user can change it.
When in doubt, Full. The cost of a full look is a few minutes; the cost of a quick look at a one-way door is the door.
Dispatch one fresh agent, not several: in Claude Code, a general-purpose
agent; elsewhere, whatever starts with an empty context. Send it four things
and nothing else: the analysis prompt copied exactly from the bottom of this
skill, MODE: Full or MODE: Quick, LANGUAGE: <the language above>, and the
brief. Never send an instruction to run this skill: the fresh session would
load it and start over.
If no fresh agent can be started, run the prompt yourself with the same MODE, LANGUAGE and brief. The prompt says you did not help make this decision; for you that is false, so work from the brief alone and attack your own reasoning hardest. Never call the result independent.
The user never sees what the agent returned: a tool result is visible only to you. So your final message is that report, copied whole, first line to last. One line goes before it: the depth, and whether a fresh agent ran (if not, name your tool and say this session wrote the report). Anything of your own comes after it, never instead of it: no summary in its place, no "details above", no reassurance the analysis did not earn, no dropped card because the user seemed committed. Bad news that arrives late is worth nothing.
Then offer to turn the "Fix before you commit" list into real things: edits to the plan, tickets, an owner and a date per item, a reminder for each early warning. "Go ahead" is delivered just as plainly.
The reasoning behind these is in references/why-these-rules.md.
The block below is the whole of
references/premortem-prompt.md, inlined
when this skill loads through scripts/prompt.sh, so no file has to be read
at dispatch time. If it shows a literal command instead, your tool does not
run inline commands: open that file next to this one and use its full text.
!bash ${CLAUDE_SKILL_DIR}/scripts/prompt.sh
>-
---name: os-what-could-go-wrongdescription: >----# skills/os-what-could-go-wrong/SKILL.mdFAQ
게시자에게서 skills/os-what-could-go-wrong/SKILL.md를 다운로드하거나 경로를 복사해 저장소를 엽니다. PlugMCP는 스킬을 목록에 올릴 뿐 파일을 호스팅하지 않습니다.
게시자가 https://github.com/kharmanskyi/open-steps에서 관리합니다.
Claude, Cursor, Codex, Grok Build는 게시자의 저장소에서 추가한 SKILL.md를 읽을 수 있습니다. Grok Bot은 채팅에서 스킬의 GitHub 폴더 링크를 보내면 설치합니다.