grounded-reply
/grounded-reply
Prowadzi bazę źródeł dla odpowiedzi na zgłoszenia klientów i ocenia, czy źródło rzeczywiście odpowiada na pytanie.
- Autor
- aomni-com ↗
- Repozytorium
- open-duet ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 30.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 30.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Prowadzi bazę źródeł dla odpowiedzi na zgłoszenia klientów i ocenia, czy źródło rzeczywiście odpowiada na pytanie.
Oryginalny opis (EN)
Build and maintain the source ledger a support reply is grounded in, and decide whether a source actually answers the ticket. Use when drafting a reply, when a source is close but not exact, when two sources disagree, or when deciding whether a question is answerable at all.
Jak działa
- One row per fact
- The coverage test
- When sources disagree
- Growing the ledger
- Before the card is written
Podgląd SKILL.md
--- name: grounded-reply description: Build and maintain the source ledger a support reply is grounded in, and decide whether a source actually answers the ticket. Use when drafting a reply, when a source is close but not exact, when two sources disagree, or when deciding whether a question is answerable at all. --- # Grounded Reply `sources.md` is not a reading list. It is the set of things you are permitted to say, and a reply is assembled *from* it rather than written and then footnoted. `AGENTS.md` says where it lives and what the ticket card looks like; this is how to keep it honest. ## One row per fact | Column | Rule | | --- | --- | | `id` | `s-001` upward. This is what a citation points at, so it never changes meaning — a rewritten source gets a new id. | | `fact` | One assertion, in the words a customer would need. If it contains "and", it is two rows. | | `where` | The URL you opened, the file path, or the resolved ticket id. Not a homepage, not a search result, not "the docs". | | `verified` | The date you last read the source itself. Not the date the source was published, and not the date you first added the row. | | `volatile` | `yes` for prices, quotas, limits, plan entitlements, setting names and URLs. `no` for concepts and procedures. Drives the staleness rule. | A resolved ticket is a legitimate source and often the best one, because it is the answer a human already gave. Cite it by ticket id and record what was resolved, never the customer it was resolved for. ## The coverage test Before writing a sentence, answer: **does this row state the customer's case, or is it merely about the topic?** This is where invented answers come from — a source that covers the neighbouring situation reads like an answer while you are holding the ticket. Three failures worth naming, because each one feels like coverage: 1. **Different plan, tier or region.** A limit documented for one plan says nothing about another. If the ticket does not say which the customer is on, the verdict is `NEEDS-INFO`, not a best guess. 2. **Different version or surface.** A settings path in the web app is not the settings path in the mobile app or the API. 3. **The general case, asked about the specific one.** "Exports are supported" does not answer "can I export 400,000 rows". If you cannot quote the row's `fact` into the reply almost unchanged, it does not cover the ticket. Escalate as `unknown-fact` and write the question into `gaps.md`. ## When sources disagree Two sources giving different answers is not a judgment call — it is a documentation defect, and the customer is not the place to resolve it. Check first, in order: are they about the same plan and surface; is one simply older (compare `verified`, and re-read the newer source rather than trusting the row); do they share an origin, so one is a copy of the other rather than a second opinion. If a real conflict survives all three, escalate as `unknown-fact` with both ids, and mark both rows `contested`. A contested row cannot ground an `ANSWER` until a human retires one of them. Never average, never pick the more recent one silently, and never pick the one that makes the reply easier to write. ## Growing the ledger The ledger is the compounding asset here — the second pass is cheaper than the first only if answers came back into it. - When a human answers an escalation, add the answer as a row the same day, `where` pointing at the escalation. That is what turns one human answer into every future reply. - When you re-read a source and it still says the same thing, update `verified`. Re-verification is cheap; a stale row that quietly grounds a price quote is not. - When a source turns out to be wrong, delete the row and add a `gaps.md` entry. Leaving it with a caveat means someone cites it later. ## Before the card is written Every factual sentence in the draft appears in the Citations table with a row id. If one does not, either find the row or delete the sentence. The reply cannot contain a claim the ledger does not — and a deleted sentence usually becomes the "Not known" line, which is the useful half of the ticket.
Wersje
W bazie nie ma innych skilli o nazwie grounded-reply.
Migawka bazy z 13.11.2026 · wersja paczki d0238f957aef (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
- 90support-reply
Pisze projekt odpowiedzi na wiadomość, zgłoszenie, opinię lub czat klienta, także gotowe szablony odpowiedzi. - 85sup-210-reply-within-policy
Pisze odpowiedź dla klienta, która odpowiada na każde pytanie, mówi tylko to, co wspiera pisemna polityka i zamówienie, i dotrzymuje wcześniejszych obietnic. - 100support-ticket-reply
Pisze empatyczne, rzetelne odpowiedzi na zgłoszenia klientów, z rozwiązaniem na początku i jednym jasnym następnym krokiem. - 90support-replies
Pisze krótkie, uprzejme i jasne odpowiedzi dla klientów na każdą wiadomość wsparcia technicznego. - 90drafting-customer-replies
Przygotowuje odpowiedź na mail klienta w sprawie rozliczeń, odnowienia lub awarii, opartą na koncie i aktualnych artykułach pomocy. - 90support-macros
Tworzy szablony odpowiedzi na powtarzalne pytania klientów, które przyspieszają pracę agentów, a jednocześnie nie brzmią jak automat.