author-plan
/author-plan
Przygotowuje lub poprawia plan.yaml projektu przez strukturalny wywiad, bez pól wyliczanych.
- Autor
- nikita-volkov ↗
- Repozytorium
- planning-dsl ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 07.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 07.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Przygotowuje lub poprawia plan.yaml projektu przez strukturalny wywiad, bez pól wyliczanych.
Oryginalny opis (EN)
Author or refine a plan.yaml for a team (humans and agents) through a structured interview. Use when the user wants to plan a project from scratch, or to edit, extend or re-estimate an existing plan.yaml. Emits only authored fields; never computed ones.
Jak działa
- Interview flow
- Hard rules
- Feeding the format back
- Validation
Podgląd SKILL.md
--- name: author-plan description: Author or refine a plan.yaml for a team (humans and agents) through a structured interview. Use when the user wants to plan a project from scratch, or to edit, extend or re-estimate an existing plan.yaml. Emits only authored fields; never computed ones. --- # Author a plan The artifact is `plan.yaml`: tasks as a dependency DAG with roles, executors (humans and agents uniformly) with capacity, the working-day calendar, and the two optional plan-level dates. It is a description of the world — nothing computed (assignment, schedule, critical path, renders) ever enters the file. The authoritative format reference is `specs/plan-format.md` in the planning-dsl repository; the canonical example at the top of that spec is the style to match. If a plan already exists, read it and validate it before asking anything: the errors and warnings tell you what the next conversation should fix. ## Interview flow Work through the stages in order; skip a stage only when the plan already answers it. Keep it light: propose, then confirm — few questions at a time, each with a recommended answer, answerable by number. Never interrogate; a proposal the person corrects is faster than a question they must answer from nothing. 1. **Frame.** What is being planned, and by when? Record `unit: person-hours`. Ask for `start:` (when execution begins) and `deadline:` — both optional: a date the person has not committed to is invented data, worse than none, so leave a field out rather than guess it. Know the consequences, not more: the scheduler refuses a plan without `start`, and omits slippage without `deadline`. 2. **Roles.** What kinds of work does this team do? Propose a short vocabulary (three to six roles) with one-line descriptions, and confirm the spellings — every task and executor will reuse these exact tokens. The descriptions are for future sessions and editors; write them plainly. 3. **Decompose.** Goals into tasks. Propose the task list and its `depends-on` edges, then grill lightly for the missing ones: what must exist before X can start? What can run in parallel with Y? A task that needs two roles is two tasks. Name each task as a slug; whatever the name wants to say goes into its `description:`. 4. **Estimate.** One person-hours figure per task. If the person hesitates or offers a range, take the middle and say you did. Leaving tasks unestimated is fine — name which ones, because scheduling stays locked until every not-done task has a number. Anything over 160 person-hours (a working month at 8 h/day) does not get estimated — it gets split: a task that size cannot be estimated honestly and the validator refuses it ([a-task-estimated-over-a-working-month-is-refused](../../../decisions/a-task-estimated-over-a-working-month-is-refused.md)). 5. **Calendar and executors.** Which days are working days? Absent `calendar:` means Mon–Fri, no holidays — if the team's week differs, or there are holidays, record `workdays`/`holidays`/`extra-workdays` as authored dates; expanding a jurisdiction's holidays into dates is this skill's job, never a `calendar: ru` string. Then who (or what) does the work? Humans: roles, plus `availability` when they are not full-time (a scalar, or sparse periods — a vacation is one `fraction: 0.0` period; days outside periods are full). Agents: roles, `concurrency` slots, and `calendar: all` — they work every day. Tracker `ids` only if they already exist in the tracker — never invent ids. 6. **Emit and validate.** Write `plan.yaml` in the spec's style, validate it, fix every error, and surface the warnings: a role no executor fills is usually a vocabulary miss, not a staffing fact. ## Hard rules - Task, executor and role names are slugs — lowercase Latin or Cyrillic letters (`ё` included) and digits, hyphen-separated, starting with a letter, one script per name (`plan-out-the-tasks`, `перегородки-первый-этаж`, not `Plan Out The Tasks` or `fix_bug`). Prose goes in `description:` (tasks) or the role's description — never in a name ([identifying-names-are-lowercase-latin-or-cyrillic-slugs](../../../decisions/identifying-names-are-lowercase-latin-or-cyrillic-slugs.md)). - Authored fields only: `unit`, `start`, `deadline`, `calendar`, `roles`; tasks (`description`, `status`, `role`, `depends-on`, `estimate`, `blocked-reason`); executors (`ids`, `roles`, `role-weights`, `availability`, `concurrency`, `calendar`). Never write `assignee`, per-task dates, schedule, per-task start/end, critical path, maturity — or `priority`, which does not exist in v1. - `status` and `blocked-reason` belong to tracker sync. One exception: when creating a plan for work already underway, seed `done` or `blocked` for what the interview establishes happened before the plan existed — then stop touching status. - A `role:` token must match the declared vocabulary exactly — `Dev` is not `dev`, and the validator will say so. - Estimates are single numbers in the plan unit. Ranges are a v2 (PERT) feature and stay out of the file. - The file describes the world as it is, not as it should be: capacity that does not exist yet is fiction the scheduler will happily schedule around. ## Feeding the format back Every point where the interview strains the schema — a shape the person wants but the format cannot say, a field that keeps feeling wrong — is a candidate change to the format design. Say so at the end of the session; it belongs in `specs/plan-format.md` or `decisions/` in the planning-dsl repository, not silently worked around in one plan. ## Validation From the planning-dsl repository root: ``` cargo run -q -- validate <path-to-plan.yaml> ``` Zero errors is the bar; warnings are worth surfacing.
Wersje
W bazie nie ma innych skilli o nazwie author-plan.
Migawka bazy z 13.11.2026 · wersja paczki 11fce2d0cf31 (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
- 90interview-me
Zamienia niejasny cel w ustrukturyzowany, żywy plan przez wywiad lub szybkie notatki i prowadzi go jako dokument. - 100graph-backbone
Definiuje, poprawia i replanuje strukturalny szkielet projektu w pliku graph.yaml w katalogu głównym. - 100plan
Prowadzi pełny cykl planu wipnote: wywiad z bramką triage tworzący karty wycinków w YAML, walidację, krytykę, przegląd człowieka, decyzje, specyfikacje i promocję. - 100plan-a-feature
Buduje specyfikację funkcji przez dociekliwy wywiad oparty na dowodach, który krok po kroku przechodzi drzewo decyzji projektowych i opisuje zachowania systemu bez szczegółów implementacji. - 95plan
Planuje zmianę: bada równolegle, pyta tylko o decyzje użytkownika i zapisuje warstwowy plan, nie pisząc kodu. - 100agenda-plan-refine
Sprawdza i poprawia istniejący plan w .agenda, aż stanie się konkretny, oparty na dowodach i gotowy do wdrożenia.