Przejdź do treści
Biblioteka skilli
AI i agenci · Przegląd i audyt

why-therefore

/why-therefore

Sprawdza spójność szkicu odpowiedzi, uzasadniając kluczowe twierdzenia i poprawiając te, które nie wytrzymują kontroli.

Licencja
MIT
Ostatnia zmiana repozytorium
13.11.2026

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 13.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

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

Co robi i kiedy

Sprawdza spójność szkicu odpowiedzi, uzasadniając kluczowe twierdzenia i poprawiając te, które nie wytrzymują kontroli.

Wyzwalacze z opisu

why?therefore?run the ritualare you sure?sanity-check thisdouble-check your reasoning
Oryginalny opis (EN)

Run the why/therefore consistency ritual on a draft answer before committing to it. For each load-bearing claim, abduce its cause ("why?"), then check blind that the cause alone leads back to the claim and not to a plausible wrong alternative ("therefore?"), and revise every claim that fails. Use this whenever the user asks to "run the ritual", asks for a why/therefore or causal-consistency check, says "are you sure?", "sanity-check this" or "double-check your reasoning", and also when an answer rests on a chain of causal or factual claims where a self-contradiction would be costly, such as a

Jak działa

  1. The steps
  2. What to show the user
  3. Example
  4. When not to use it
  5. Limits
Sekcje z nagłówków SKILL.md, w oryginale · 98 linii, 1 158 słów
Podgląd SKILL.md
---
name: why-therefore
description: Run the why/therefore consistency ritual on a draft answer before committing to it. For each load-bearing claim, abduce its cause ("why?"), then check blind that the cause alone leads back to the claim and not to a plausible wrong alternative ("therefore?"), and revise every claim that fails. Use this whenever the user asks to "run the ritual", asks for a why/therefore or causal-consistency check, says "are you sure?", "sanity-check this" or "double-check your reasoning", and also when an answer rests on a chain of causal or factual claims where a self-contradiction would be costly, such as a root-cause analysis, a diagnosis, an explanation of why something happened, a plan with dependencies, or a summary of evidence.
---

# Why/Therefore ritual

"Why" is the inverse of "therefore". If a claim is something you really hold, it must survive a round trip: from the claim to its cause, and from the cause back to the claim. A claim that cannot make the trip back is a claim you wrote, not a claim you can support.

This skill is the in-context form of a training idea (a cycle-consistency loss between "why" and "therefore"). Nothing is trained here. The cycle runs on your draft answer, and each failure makes you revise the draft.

Status: this ritual is untested. No one has measured that it improves answers. Treat it as a structured self-check, not as proof.

## The steps

### 1. Draft, then list the claims

Write your answer as a draft. Do not send it. List the load-bearing claims, one sentence each. A claim is load-bearing if the answer changes when the claim is false. Three to seven claims is the usual number. More than that usually means the claims are too small.

### 2. Why: abduce a cause for each claim

For each claim, write one sentence that says why it is true.

Do not reuse the content words of the claim in the cause. If the cause repeats the claim, step 4 becomes a word match and tests nothing. "The build fails because the build is broken" will pass every check and tell you nothing.

### 3. Write a hard distractor for each claim

A distractor is the answer that a careless version of you could have given. It has the same topic and the same form as the claim, but different substance: a different culprit, the opposite direction, a different number.

An easy distractor makes the test worthless. "The cache is stale" against "Tim got a promotion" is not a test. "The cache is stale" against "The cache key is wrong" is a test.

### 4. Therefore: go back from the cause, blind

The question is: given only the cause, which follows, the claim or the distractor?

The reader of the cause must not know which option is the original. You wrote the claim, so you cannot do this step honestly with the claim in view.

- If you can start a subagent, use one for each claim, or one for the full batch. Give it only the cause and the two options in a random order, labelled A and B. Ask: "Which statement follows from this cause: A, B, or neither? Answer with one letter and one line of reasoning." Do not give it the user's question, your draft, or a hint.
- If you cannot start a subagent, do the step yourself as honestly as you can. Read only the cause line. Write what follows from it before you look at the options. Then compare.

### 5. Read the verdicts

- **Pass:** the reader picked the claim.
- **Fail:** the reader picked the distractor, or "neither".

A fail means the cause does not support the claim. There are two possible reasons. The cause is wrong, so go back to step 2 and find the real one. Or the claim is wrong or too strong, so change it, hedge it, or remove it.

Do not reword the cause until the reader is forced to pick the claim. That makes the ritual agree with itself and teaches you nothing. If you cannot find a cause that leads back to the claim without reusing its words, the claim is the problem.

### 6. Go to the root

For each claim that passed, ask "why?" again on its cause. Stop after three levels, or when you reach bedrock. Bedrock is something given: a fact from the user, a file or a source that you read, a definition.

Then compare the roots. Chains about the same subject must end in roots that can all be true at the same time. Two roots that contradict each other mean that the answer holds two views of the world. Resolve the contradiction before you answer.

A chain that never reaches bedrock is a claim with no support in this conversation. Say so in the answer, or go and get the evidence.

### 7. Paraphrase probe

Restate the user's question in two other ways. If your answer changes with the wording, the answer depends on the phrasing and not on the facts. Find out which reading the user means, or give an answer that holds for all of them.

### 8. Rewrite

Write the final answer from the claims that survived. Evidence in the conversation (files, tool results, sources, facts from the user) always outranks the result of the cycle. The cycle checks that you agree with yourself. It does not check that you are right.

## What to show the user

By default, show only the final answer and one line that says what the ritual changed, for example: "Check: 5 claims, 1 revised (the cache claim had no cause that led back to it)."

If the user asked for the ritual by name, also show the ledger:

| # | Claim | Why (cause) | Distractor | Verdict | Action |
|---|---|---|---|---|---|
| 1 | ... | ... | ... | pass / fail | kept / revised / removed |

Then list the roots from step 6 and any contradiction you found.

## Example

Question: "Why did the plants in my garden wilt?" The draft says: "They wilted from lack of water."

- Claim: The plants wilted from lack of water.
- Why: There has been no rain for three weeks and the hose was not used.
- Distractor: The plants wilted from too much fertiliser.
- Therefore, blind: from "no rain for three weeks and the hose was not used", the reader picks the claim. Pass.
- Root: "no rain for three weeks" is a fact from the user. Bedrock.

Now suppose the cause had been "Plants need the right conditions to grow". From that cause alone, the reader cannot choose between the claim and the distractor. That is a fail: the draft had a conclusion but no reason behind it. The fix is to ask the user about watering and fertiliser, not to assert the first guess.

## When not to use it

- Simple lookups and one-step answers. The ritual costs more than a mistake does.
- Creative writing, brainstorming and opinion, where there is no claim to check.
- Pure mathematics or code behaviour. A proof check or running the code is a better test than a cycle of sentences.

## Limits

- Untested. See the status note at the top.
- Consistent is not the same as correct. A wrong belief that is held consistently passes every step.
- If you do step 4 yourself, you know the answer already, and the test is weaker.
- The ritual adds model calls and time. Use the batch form of step 4 when there are many claims.

Wersje

W bazie nie ma innych skilli o nazwie why-therefore.

Migawka bazy z 13.11.2026 · wersja paczki 87c9b4cd0281 (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

  • 90opus-verify
    Sprawdza każde twierdzenie lub zmianę w pracy względem dowodów, takich jak testy, wyszukiwanie w kodzie czy odczyt pliku, i oznacza je jako potwierdzone, błędne lub niesprawdzone.
  • 100verify-claims
    Uruchamia Chain-of-Verification (CoVe) na szkicu z twierdzeniami faktograficznymi przez agenta claim-verifier w świeżym kontekście i raportuje, które są poparte, zaprzeczone lub niesprawdzalne.
  • 100double-check
    Weryfikuje wcześniejszą odpowiedź na prośbę o sprawdzenie lub powtórzenie analizy, wyprowadzając wynik ponownie ze źródła.
  • 100double-check
    Weryfikuje wcześniejszą odpowiedź, gdy użytkownik pyta „czy na pewno”, i powtarza analizę od źródła, żeby sprawdzić wynik.
  • 100self-consistency
    Generuje N niezależnych ścieżek rozumowania i głosuje nad odpowiedzią, przy kompromisach architektonicznych i niejednoznacznych decyzjach projektowych (Wang i in., 2022).
  • 90answer-audit
    Audytuje odpowiedź systemu RAG wobec źródeł, zakładając błąd, dopóki każde twierdzenie nie zostanie potwierdzone dosłownie.

Uwagi i zgłoszenia

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