Przejdź do treści
Biblioteka skilli
AI i agenci · Automatyzacja

brief

/brief

Deleguje jedną jednostkę pracy jednemu nazwanemu agentowi: wypełnia pojedynczy MANDATE, wysyła go i milknie.

Repozytorium
agent-skills ↗
Licencja
MIT
Ostatnia zmiana repozytorium
27.10.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 27.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Deleguje jedną jednostkę pracy jednemu nazwanemu agentowi: wypełnia pojedynczy MANDATE, wysyła go i milknie.

Argumenty

/brief [work description]
Oryginalny opis (EN)

Delegate one unit of work to one named agent: fill a single MANDATE, send it, then go silent. Use when: brief, delegate, mandate, BRIEF state, delegate to a named agent, hand a unit of work off, put an idle agent to work, send a mandate.

Jak działa

  1. 1. Source the work, and the boundary
  2. 2. One message to the human
  3. 3. Fill
  4. 4. Append these five lines verbatim, and send
Sekcje z nagłówków SKILL.md, w oryginale · 62 linii, 915 słów
Podgląd SKILL.md
---
name: brief
description: 'Delegate one unit of work to one named agent: fill a single MANDATE, send it, then go silent. Use when: brief, delegate, mandate, BRIEF state, delegate to a named agent, hand a unit of work off, put an idle agent to work, send a mandate.'
allowed-tools: [SendMessage]
argument-hint: "[work description]"
context: main
user-invocable: true
---

# Brief

BRIEF emits **one** mandate, to **one** named agent, then you are in WATCH — silent until a question or a report comes back. Never broadcast. Runs in the originating window: usually the orchestrator, but any agent that hands work off.

**Read nothing here.** Skeleton and role lines below are complete and verbatim — they are the whole contract the delegate needs, and you must assume it has read nothing else. `../../references/protocol.md` holds the reasoning behind each field, for whoever amends the contract; opening it in this window spends the context this skill exists to protect.

**Choose no model.** The delegate is a live agent in a window that already exists — its model was fixed when that window was opened, and a mandate cannot change it.

## 1. Source the work, and the boundary

Arguments describe the work → `scope` and `knowledge_state`. Bare → ask. Never infer from the surrounding conversation, never invent.

`owned_paths` / `hands_off` you assert from what you already hold. No plan file to consult, no roster to inherit. This is the one thing that cannot be deferred: two agents writing one path lose updates with no error at all — silent, unrepairable.

## 2. One message to the human

Fold every outstanding question into a single plain-text message: the target's **name**, plus whatever step 1 left unsourced.

## 3. Fill

```
scope:               what is yours / what is explicitly not yours
orchestrator:        the name of the agent sending this mandate — your only machine recipient
owned_paths:         [ABSOLUTE paths; trailing / = whole subtree — outside them you read, you do not write]
hands_off:           [shared, high-blast-radius files — a reminder on top of that rule, never the boundary]
read_first:          [ABSOLUTE paths, in order — a closed list. Anything else, ask me before opening it]
deliverable:         ABSOLUTE path you own — your findings land there, not in your report
knowledge_state:     established / hypothesis / to_discover
stop_conditions:     ... (must include "you discover the scope is wrong")
cadence:             when you are expected to report
what_i_do_not_know:  ...
```

- `scope` — state the negative half. An unstated boundary is one an agent crosses.
- `orchestrator` — **your own name**: the delegate's only machine address, where its REPORT lands. Omitted → reporting becomes something only the human can relay by hand. One name is not a roster; questions and readbacks still print in the delegate's own window, for the human.
- `owned_paths` — absolute, so nothing resolves against the delegate's cwd. Precise file when it exists, trailing `/` for a subtree it will populate. A subtree is owned **exclusively**: "you both write in `X/`, just use different filenames" partitions nothing — two agents asked to write up their notes both create `notes.md`.
- `hands_off` — a few shared, high-blast-radius files it is likely to reach for. A reminder, never the boundary; the boundary is default deny.
- `read_first` — **closed, ordered, and yours to decide.** `owned_paths` partitions writing; this partitions the delegate's incoming context, which is otherwise the one thing it spends entirely on its own judgement. Left empty, an agent handed a plan, an audit and a journal opens all three and a few neighbours "to be sure", and saturates before its first useful move. Name only what the work needs, and close the list: *anything else, ask me before opening it*. This does not contradict "read nothing at runtime" — that rule is about the contract, never about the project's own documents.
- `deliverable` — one absolute path, inside `owned_paths`, where its findings accumulate as it works. The REPORT then carries ten lines and this pointer instead of the content. Without it every fact lives only in a message: it inflates your window on arrival, and dies with the delegate's.
- `stop_conditions` — always includes "you discover the scope is wrong".
- `cadence` — criterion is **reversibility**, not milestones, not elapsed time. Name the points of no return where it stops and waits; everywhere else it reports and keeps going while you redirect asynchronously. Halting at every milestone makes the human the bottleneck.
- `what_i_do_not_know` — mandatory, never empty. Hidden gaps get filled by plausibility at the other end.

## 4. Append these five lines verbatim, and send

```
Announce your state on a single line at every change: [READY] [READBACK] [WORK] [QUESTION] [REPORT].
In READBACK: three lines — mission as I understand it / assumptions I filled in / blocking gaps — then wait for my confirmation.
If you discover that the scope you were given is wrong, switch to QUESTION immediately. That takes priority over the task.
You emit a message only on a state transition. The rest of the time you are silent.
When a stop condition or a cadence checkpoint fires, run team:report. It is the only message you send by machine.
```

Every mandate carries them. They are the delegate's entire contract — states, readback, the wrong-scope override, and the rule that it speaks only on a transition. Filled skeleton and role lines go out as **one** `SendMessage`, never two.

**If the send fails** — tool error = delivery failure, not a state transition. Name it to the human in plain text and stop. No retry loop, no neighbouring name: a silent retry against a dead name loses a whole branch of work, a near-miss name drops this mandate in someone else's window.

---

One mandate, one named target, then WATCH. The target answers in READBACK — in its own window, to the human, not back to you.

Wersje

RepozytoriumWynikLicencja
Ta strona · digital-stoic-org/agent-skills100MIT
sekretmaksym-dev/design-skills100MIT
wrg32786/aigent-os100MIT
magnifito/domain-name-engine100MIT
alirezarezvani/claude-skills100MIT
anthropics/knowledge-work-plugins100Apache-2.0
educlopez/ui-craft100MIT
oliver-kriska/claude-elixir-phoenix100MIT
hirogakatageri/hirokata-claude-marketplace100MIT

Migawka bazy z 13.11.2026 · wersja paczki 4de50dcdb07a (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 100ask
    Deleguje zadania lub prosi o informacje innego agenta zarządzanego przez CCB poleceniem ask, na przykład przy przekazaniu pracy wskazanemu agentowi.
  • 90handoff
    Pisze krótką instrukcję i przekazuje jedno zadanie agentowi Claude Code, Codex lub Antigravity, albo zapisuje przekazanie.
  • 90delegate
    Pomaga oddać samodzielny kawałek pracy innemu agentowi: wybrać, co wydzielić, napisać zrozumiały opis, sprawdzić wynik.
  • 100agent-brief
    Pokazuje, jak pisać zlecenia dla agentów pomocniczych i ile naraz uruchamiać równolegle.
  • 100delegate
    Deleguje implementację do Cursora przez narzędzie delegate z cursor-delegate-mcp: planowanie przed budową, pytania o bazę kodu, wznawianie sesji delegacji i diagnozowanie niedziałającej delegacji.
  • 100a2a-cli
    Deleguje zadania agentom A2A z wiersza poleceń: pobiera kartę agenta, wysyła wiadomość i sprawdza lub anuluje zadanie.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.