start-unit
/start-unit
Rozpoczyna jednostkę dekompozycji: ładuje specyfikację, weryfikuje założenia, pracuje do kryteriów ukończenia i przechodzi przez wdrożenie, przegląd i triaż.
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Rozpoczyna jednostkę dekompozycji: ładuje specyfikację, weryfikuje założenia, pracuje do kryteriów ukończenia i przechodzi przez wdrożenie, przegląd i triaż.
Oryginalny opis (EN)
Begin work on a decomposition unit. Loads the spine and the unit's spec, reads the code the unit touches, verifies every assumption and records the boundary in docs/decisions.md, works the unit to its done-when criteria, then chains through ship, review, and triage, stopping only on a rule-6 stop condition or at the merge gate. Use when starting a new unit of Olympus v2 work. Takes a unit id, e.g. start-unit P1.
Jak działa
- 1. Load context, in this order
- 2. Check dependencies
- 3. Read the code the unit touches
- 4. Verify the boundary and every assumption, record it, then proceed
- 5. Work the unit
- 6. Stop conditions
- 7. Finish
Podgląd SKILL.md
--- name: start-unit description: Begin work on a decomposition unit. Loads the spine and the unit's spec, reads the code the unit touches, verifies every assumption and records the boundary in docs/decisions.md, works the unit to its done-when criteria, then chains through ship, review, and triage, stopping only on a rule-6 stop condition or at the merge gate. Use when starting a new unit of Olympus v2 work. Takes a unit id, e.g. start-unit P1. --- # Start a unit This replaces a hand-written kickoff prompt. Everything a session needs lives in the unit spec; this skill loads it and holds the boundary. ## 1. Load context, in this order 1. `CLAUDE.md` — invariants, vocabulary, conventions 2. `docs/plan/F1-spine.md` — invariants, frozen vocabulary, dependency graph 3. `docs/plan/WORKFLOW.md` — the three work types, the unit loop, where state lives 4. `docs/plan/DECOMPOSITION.md` — the named unit's entry only 5. The unit's detailed spec if one exists, e.g. `docs/plan/F2-contracts.md` Load nothing else. The maintainer's master plan is not in the repository and is never session context; `CLAUDE.md` says where it lives and that it stays out. ## 2. Check dependencies The unit entry names what it depends on. Verify each dependency has shipped — its branch merged into `v2`, its acceptance criteria met. If one has not, say so and stop. A unit built on an unshipped dependency will be rewritten. ## 3. Read the code the unit touches The unit entry is a claim about the code, not evidence of it. Read what the unit will change or build on: the packages it names, the types it consumes, and the pending registry entries it owes. Note every place where the code differs from what the entry says. An entry that says an input already exists has been wrong before. ## 4. Verify the boundary and every assumption, record it, then proceed Write the following as the unit's first entry in `docs/decisions.md`, print it, and continue to step 5 without waiting (`CLAUDE.md`, working rule 6). The entry ships in the review bundle, so both reviewing families challenge it; that replaces a confirmation typed by the maintainer, and it is only as good as the verification behind each line. - Unit id and title - Deliverables, one line each - **Out of scope**, verbatim - Acceptance criteria, as the commands that will prove them - Which invariants this unit touches, and which pending registry entries it owes - **Every assumption the plan rests on**, each verified, naming the file:line or the command that verified it. Run what can be run rather than reason about it. An assumption that cannot be verified from here is a stop condition (step 6), not a guess - Where the code differs from the unit entry, from step 3 - Every choice point the spec leaves open: the options, each checked against the code, the one chosen, what was rejected and why, and how to reverse it. Choose the strongest on correctness, security, and the invariants — never the easiest, and never the one that defers work to a later unit without an owner If any of the first five is absent from the spec, propose what it should be. Where the proposal follows from the spine, the dependency graph, and the neighbouring units' entries, fix `docs/plan/DECOMPOSITION.md` in the unit's first commit and record it here. Where it does not — the missing piece would decide the unit's scope or what proves it done — stop and ask, with the proposal as the recommended answer. ## 5. Work the unit - Conformance suite first. It is the acceptance criteria in executable form. - Stay inside scope. Adjacent work becomes a note in the PR body, not a commit. - Record every judgment call in `docs/decisions.md`: what was ambiguous, what was chosen, why, how to reverse it. - Invariants outrank convenience. A change that makes the unit easier and weakens I1–I10 is wrong, however clean. - Pay down pending registry entries this unit owns. Adding a new one requires editing the committed baseline — a deliberate act, visible in the diff. ## 6. Stop conditions Stop and ask when: - The spec contradicts itself, or contradicts `F1-spine.md` - Meeting an acceptance criterion would require weakening an invariant - The work needs a contract change (those are amendments, not unit work), or a gate-path change the unit does not declare in scope - A destructive operation has no safe default - A discovery would change the unit's scope, deliverables, or acceptance criteria - The next step depends on something that has not been verified and cannot be verified from here Everything else is decided without waiting: verify the options against the code, take the strongest, record it in `docs/decisions.md` with what was rejected and how to reverse it, and continue. A fact that turns out different from step 4's record is corrected in that record, in a commit of its own, and the work continues unless the correction meets a condition above. When you stop, ask every open question together, each with its recommended answer, and nothing else. ## 7. Finish Run the acceptance criteria. When they pass, invoke `ship-unit`. From there the chain runs `review-request`, `run-review`, and `triage-review` without stopping to ask, and ends at the merge gate (`docs/plan/WORKFLOW.md`, step 7). Do not begin the next unit. Every unit boundary is a human review gate.
Wersje
W bazie nie ma innych skilli o nazwie start-unit.
Migawka bazy z 13.11.2026 · wersja paczki 460ad91820d6 (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
- 100spec-unit
Pisze specyfikację jednego etapu budowy na podstawie kontekstu projektu i planu, pytając o niejasne decyzje. - 100work-unit-protocol
Wymusza kontrakt raportowania ULTRA §13 przy rozpoczynaniu, śledzeniu i zamykaniu jednostki pracy: co, dlaczego, testy, zależności, ponowne użycie, ryzyka, następne kroki i tabela ukończenia. - 100ship-unit
Otwiera PR dla ukończonej jednostki pracy, sprawdza kryteria akceptacji i prosi o zewnętrzną recenzję. - 100update-unit
Podnosi jednostkę do najnowszej wersji upstream i sprawdza ją testową budową. - 100new-unit
Tworzy lub poprawia jednostkę nawyków dla języka, narzędzia lub praktyki testowej: wyzwalacze, treść, rejestrację i walidację. - 90build-unit
Realizuje jedną jednostkę zaplanowanej budowy dokładnie według pliku specyfikacji, sprawdza ją na liście kontrolnej i aktualizuje tracker.