메뉴
Pause, surface the current plan + open questions, and suggest the next concrete step so the agent (or operator) can recover from a deadlock or analysis paralysis.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | Get Unstuck |
| description | Pause, surface the current plan + open questions, and suggest the next concrete step so the agent (or operator) can recover from a deadlock or analysis paralysis. |
| category | meta |
| effort | low |
| when_to_use | When the conversation has been spinning — repeated tool calls without progress, contradictory edits, the agent admitting it doesn't know what to do next, or the operator typing "you're stuck" / "let's try again". Don't use this as a first response — use it when the rope's already in a knot. |
| version | 1.0.0 |
Stop. Don't make another tool call. Read this whole skill before deciding what to do next.
You probably got here because one of these is true:
If none of those describe the moment, this skill probably isn't what you needed.
What was the user's original ask? Not your interpretation, not the third-derivative subgoal you ended up chasing. The literal prompt or its plain-English rephrasing.
Not what you tried, not what you intended. What's committed to disk or visible in the conversation right now. Bullet list of ≤ 5 items.
What single piece of information would unblock you? It's almost always one of:
Pick the smallest action that produces new information. Not a plan, not a refactor — one Read, one Bash command, one user question. The goal is to break the symmetry of being stuck.
End with a single yes/no question to the user. Examples:
Make the question answerable with one short reply. The point is to put the user back in the driver's seat without dumping your state on them.
Discard the rest of your prior plan. Treat the user's answer as the new starting point. The whole point of this skill is to reset, not to continue.
How to build a NEW tool of your own by writing its code in your sandbox, testing it, and forging it into a live, callable tool with `env(action="forge_tool")`. Use when no existing tool does what you need and you can implement it as a small script. Multiple tools + skills can be saved together as one reusable Sandbox Tool Pack.
How to inspect and edit your OWN operating environment at runtime — system prompt, active tools, and skills — using the `env` tool. Use when you need to change your own capabilities (turn a tool on/off, rewrite your instructions, author a new skill) or save them for this session.
Run the same task on a recurring cron schedule. Parses interval syntax (5m, 2h, 1d, "*/15 * * * *"), records the schedule, and acknowledges back to the user.
Review code in the current working tree across three independent dimensions (reuse, quality, efficiency) and synthesise a small punch-list of high-value cleanups.
Interview the user about a workflow they keep repeating, then generate a reusable SKILL.md and write it to disk so the next session starts with the skill already loaded.
Apply the same prompt or operation to a list of items in turn, collecting results. Useful for "do X to each of these files / PRs / records" tasks.
