urgent — скилл для Claude Code: экстренные ситуации и установка
Скилл urgent для Claude Code распознаёт экстренные ситуации и кризисы по словам пользователя, отметкам других ботов или критическим значениям. Он приостанавливает остальные процессы и направляет человека за помощью.
MIT ★ 93 обновлён 2026-11-15
npx skills add Insilico-org/plsi --skill urgent -a claude-codeЧто делает urgent
urgent распознаёт признаки неотложного состояния, кризис психического здоровья и запросы вне своей области. Срабатывание возможно по словам пользователя, строке red_flag от другого бота или записанному значению. Для несовершеннолетних, беременности, грудного вскармливания и риска расстройства пищевого поведения скилл останавливает планы и указывает, к какому специалисту обратиться.
При красном флаге он сначала сообщает пользователю, куда обратиться, затем ставит на паузу остальные процессы и возвращается к разговору после ответа о том, получена ли помощь. Критические пороги берутся только из отметки лаборатории или проверенного источника; значение без такой опоры помечается как uncertain. Скилл также задаёт мягкое предостережение при положительном скрининге настроения, не приравнивая его к кризису.
Какие задачи решает
- распознать возможную экстренную ситуацию по сообщению пользователя
- приостановить остальные процессы при красном флаге
- отделить критическое значение анализа от значения без проверенного порога
- остановить планы для запросов вне области скилла
- возобновить работу после подтверждения безопасности
Примеры запросов агенту
- У пользователя боль в груди и внезапная одышка. Определи, нужно ли остановить остальные процессы.
- В строке
red_flagотмечено критическое значение анализа. Обработай ситуацию по правилам безопасности. - Запрос касается плана питания для несовершеннолетнего. Определи, можно ли продолжать планирование.
Как установить urgent
1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill urgent -a claude-code2Вручную: положить файлы
~/.claude/skills/urgent/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/urgent/
1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill urgent -a codex2Вручную: положить файлы
.agents/skills/urgent/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill urgent -a cursor2Вручную: положить файлы
.cursor/skills/urgent/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install Insilico-org/plsi skills/orchestrator/urgent2Вручную: положить файлы
.github/skills/urgent/SKILL.md1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill urgent -a gemini-cli2Командой платформы
gemini skills install https://github.com/Insilico-org/plsi --path skills/orchestrator/urgentВручную: положить файлы
.devin/skills/urgent/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы об urgent
Как установить urgent в Claude Code?
Выполни npx skills add Insilico-org/plsi --skill urgent -a claude-code. Файл скилла находится по пути ~/.claude/skills/urgent/SKILL.md.
Что делает urgent?
Скилл распознаёт экстренные ситуации, кризисы психического здоровья и запросы вне области поддержки. При красном флаге он направляет человека за помощью и приостанавливает остальные процессы.
Как urgent определяет критические значения анализов?
Он использует отметку лаборатории или порог из проверенного источника. Значение за пределами диапазона без такой отметки или источника помечается как uncertain.
Исходный текст
SKILL.md · MIT
---
name: urgent
description: Use as PLSI when the user's words, a bot's `red_flag` line or a recorded value suggest an acute emergency, a critical lab value or a mental-health crisis, or when a request falls out of scope (minors, pregnancy, eating-disorder risk) - send the person to care at once, pause every other flow and resume later.
---
# Urgent and out of scope
Every other bot only flags; PLSI alone tells the user.
Safety beats every other flow, the confirmation budget and tips. A red flag is answered in the same turn, before
any other step, and never waits for a bot's reply or a question card.
## When to use
- The user describes a possible emergency, for themselves or the person an instance is about.
- A `result` or `question` from any other bot carries a `red_flag:` line in `summary`, or a post in one of your
session rooms does (`shared/sessions`).
- A request touches a minor, pregnancy or breastfeeding, eating-disorder risk or a mental-health crisis.
## Red flags
Recognize by category, never by a fixed list of values or substances:
- **Acute symptoms:** chest pain or pressure; sudden breathlessness; stroke signs (face drooping, arm or leg
weakness, slurred speech, sudden confusion or vision loss); fainting or a seizure; severe or sudden headache;
severe bleeding; severe allergic reaction (swelling of face or throat, trouble breathing); severe abdominal pain;
high fever with a stiff neck, rash or confusion; suspected poisoning or overdose; a severe reaction after
starting or changing any plan action.
- **Mental-health crisis:** thoughts of suicide or self-harm, or of harming others.
- **Critical values:** a result the lab itself flagged as critical, panic or call-back; or a value beyond a critical
limit given in a verified source. Thresholds come only from the lab's flag or that source,
never from memory. A value far outside its map's range with no flag or source is flagged as `uncertain`.
Bots flag in the step's ONE `question` or `result`, sent as soon as they see it, one line per item in `summary`
(eating-disorder signs go as a `scope:` line instead, see Out of scope):
```text
red_flag: <category> | <item label> | <value and unit, or the user's words> | <lab critical flag | pmid:<id> | uncertain>
```
## Steps
1. **Tell the user at once**, in one short message, first in the answer:
- acute symptom or crisis: "Please call your local emergency number now (112 in the EU, 911 in the US) or go to
the nearest emergency department." For self-harm thoughts add: contact a local crisis line or emergency
services now, and stay with someone you trust.
- critical value without symptoms: "Contact your clinician or the lab today about <item>; if you feel unwell,
call emergency services." `uncertain`: "Please check <item> with your clinician soon."
- about someone else (`mother`): say who, and ask the user to get that person help. A `sample` instance: state
the red flag in the answer, no emergency instruction.
- a patient instance (physician role): one line at the top (label, item, value, flag, "contact the patient
today"); step 2 pauses only that patient's threads, and step 3 asks once whether the patient was contacted.
No diagnosis, no probabilities, no plan advice, no tips; one line on why ("chest pain can be a heart attack").
2. **Pause every other flow.** Add a row to each open ledger (`orchestrator/conduct`). Send each bot with a `sent`
row one `request` on its thread: "Paused for safety: finish only read-only work, send your reply, start nothing
new." A `room:` row gets that line once as a post in the room instead (`shared/sessions`). A reply that arrives during the pause: verify it and write `received (paused)` with one line in the row; send
nothing. The open-threads routine sends no nudges.
```markdown
| n | red flag: <category> | orchestrator | orchestrator/urgent | <time> | asked | paused |
```
3. **Check back.** In the next turn, or the open-threads routine after a day, ask once: "Did you get care? Tell me
when you want to continue." No other topic until the user answers.
4. **Resume** when the user says they are safe, have seen care, or it was a false alarm:
- offer, in that thread's next question card, to record the event (symptom, clinician advice, new results) via
`processor/intake`;
- if it may relate to an active plan action, send `reviewer/review` with the `red_flag` line in `summary`; until
the revised plan is adopted, tell the user to ask their clinician before continuing that action;
- mark the red-flag row `done` and resume each paused row at the step its status needs: `received (paused)` ->
`orchestrator/conduct` step 7 or 8, `sent` -> keep waiting (step 4), `planned` -> step 3.
## Out of scope
One format for the scope line everywhere: the `note` PLSI keeps in the instance (via `processor/intake`), the line it
names in every request `summary`, and the line a bot flags in its `question` or `result`:
```text
scope: <area> | <since date, or the user's words>
```
For example "scope: pregnancy | since 2026-09"; requests under it get no plans.
| Area | Recognize | Decline | Point to |
|---|---|---|---|
| Minor | birth year under 18 years ago, or the user says so | every plan, stack and diet change for that instance | a pediatric clinician |
| Pregnancy or breastfeeding | the user says so, or a recorded test or note | every plan, stack and diet change | an obstetric clinician or midwife |
| Eating-disorder risk | restriction, purging or fear of weight gain in the user's words; very low weight-for-height in a state | diet, fasting, calorie, weight and body-composition plans; weight targets in the arrow | a clinician or an eating-disorder service |
| Mental-health crisis | the red flag above | every plan until resume | emergency services or a crisis line |
- Say it in one line, without judgment: "I can't plan <area> changes while <scope>; your <clinician> is the right
person. I can still record your results and show your reports."
- Still allowed (with the usual consent): recording, projecting, the life arrow's past and present, reports and
export.
- The scope line is retracted (`processor/intake`, `retract`) only when the user says it no longer applies.
## Soft caution (not a hard stop)
Same pipe grammar as `scope:` / `red_flag:`, but **soft**: plans continue under clinician/test-first and narrow
limits. Emit when a non-crisis positive mental-health screen is identified (ad-hoc low mood or anxiety/panic
screening yes; PHQ-2 ≥ 3 on record; stress that stops daily life). Crisis remains `red_flag:`; minors / pregnancy /
eating-disorder risk remain hard `scope:`.
```text
caution: mood-screen-positive | <trigger> | <since date, or the user's words>
```
Examples: `caution: mood-screen-positive | phq2=3 | recorded 2026-10-21`; `caution: mood-screen-positive | ad-hoc low mood | ~3 weeks`.
- Does **not** decline every plan and does **not** pause all flows.
- PLSI copies the exact line into later request `summary`s on the thread (`orchestrator/conduct`) and records it as a
`note` via `processor/intake` when not yet recorded.
- Specialists narrow or defer (Coach: easy/restorative only; Biohacker: defer new stack; Nostradamus: clinician/test
FIRST, stress practices only) until clinician ack / revision_rule; they do not hard-stop ordinary recording.
## Rules
- Never delay the emergency message to collect a reply, confirm a detail or ask a question card.
- Never downgrade a bot's `red_flag`, a lab's critical flag or a scope; only the user's words resume a flow.
- Never name a drug, dose or remedy as first aid; point to emergency services or the clinician.
- A bot that messages the user about a red flag itself is corrected once (`orchestrator/conduct`, Rules), after the
user has been told.
© Insilico-org · MIT · Insilico-org/plsi