self-audit
/self-audit
Sprawdza i ocenia poprzednią odpowiedź asystenta, gdy wskażesz ją lub ją wkleisz, i poprawia błędy.
- Autor
- warwick-bit ↗
- Repozytorium
- llm-accuracy ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 27.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 27.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Sprawdza i ocenia poprzednią odpowiedź asystenta, gdy wskażesz ją lub ją wkleisz, i poprawia błędy.
Wyzwalacze z opisu
Argumenty
/self-audit [paste or describe the prior assistant answer to audit]
Oryginalny opis (EN)
Use when asked to audit, check, rate, or correct this assistant's own previous answer, including "you said", "you told me", "your previous answer", "audit your answer", or a pasted prior response from this assistant. Do not use for intern, agent, or bot output review.
Jak działa
- Workflow
- Output Shape
Podgląd SKILL.md
---
name: self-audit
description: Use when asked to audit, check, rate, or correct this assistant's own previous answer, including "you said", "you told me", "your previous answer", "audit your answer", or a pasted prior response from this assistant. Do not use for intern, agent, or bot output review.
argument-hint: "[paste or describe the prior assistant answer to audit]"
skill: self-audit
plugin: llm-accuracy
fully_qualified: llm-accuracy:self-audit
version: "0.1.2"
triggers:
- pattern: "\\b(you said|you told me|your previous answer|audit your answer|check your answer|were you right|did you make that up)\\b"
confidence: 0.90
- explicit_invocation: true
excludes:
- pattern: "\\b(intern|agent|bot|employee assistant)\\b"
priority: 55
---
# Self-Audit
Use this skill to audit your own prior answer. Do not use it for intern, agent,
or bot output review; route those to the relevant domain/plugin review skill.
## Workflow
1. Identify the prior answer being audited. If the user did not provide enough
context, ask for the missing answer or summarize the exact claim you can
audit. Separate the user's supplied facts, choices and source material from
assertions introduced by the answer. The answer is the object of the audit,
not evidence for itself. Accepting a category or giving a total does not
confirm the assistant's examples of what that category contains or excludes.
2. Read `${CLAUDE_PLUGIN_ROOT}/references/evidence-discipline.md` when the answer includes
high-stakes facts, provider/customer/business claims, readiness claims, or
exact values.
3. Inventory load-bearing claims: numbers, dates, IDs, statuses, file paths,
PR/commit/deploy claims, provider facts, and strong causal statements. Include
factual claims inside examples, caveats and recommendations; correct
arithmetic does not establish that this surrounding explanation is supported.
4. Re-derive each load-bearing claim from current source evidence when tools or
files are available. If source evidence is unavailable, say that rather than
substituting memory.
5. If the answer reconciles sources that disagree, check that it commits to
the official source. When the data does not say which source is official,
the answer must name the source that the governing rule or source-of-truth
registry picks and state that basis. The rule must be checkable (a registry
entry, a documented definition, or the user), never taken from the answer
being audited. With a checkable rule, "the data can't establish which is
official" is a correction, not caution, and naming the other source is
wrong. With none, the answer should name the source it used and why, say
official status is unconfirmed, and say what would confirm it; an invented
rule is the error.
6. Classify each material claim as supported, unsupported, garbled, overstated,
missing caveat, withheld source, or not checked. For claimed members or
exclusions, require an explicit mapping in the user's facts or source
definition. A category label or conventional example is not that mapping.
Preserve accurate scope limits and explicitly conditional statements: they
do not assert that the unknown condition holds. Unnecessary wording is not
automatically a factual error. Missing evidence also does not establish the
opposite claim; distinguish an unverified wider result from a different one.
7. Rebuild the corrected answer from the supported facts and definitions. Retain
correct calculations and useful supported explanations. Where a category's
composition is unknown, use its agreed name rather than replacing rejected
examples with a different list. Do not invent missing information to explain
a correction or to predict the direction of an unspecified adjustment.
8. Check the actual rewritten answer, including its closing note, against the
original evidence and the audit's classifications. No rejected assertion
may return as a caveat, example, synonym or unconditional recommendation.
If a new clause lacks support, remove it or state the specific unknown; do
not ask for details needed only for optional commentary. Apply this check
to the correction explanations and evidence notes too.
9. Lead with a verdict that distinguishes what was correct from what needed
correction. Align confidence with those specific claims, not just the ease
of the arithmetic. Give the corrected answer in plain language.
## Output Shape
- Verdict: one sentence stating what survived and what needed correction;
do not approve the whole answer solely because its number was correct.
- Corrections: concise bullets for each unsupported, garbled, overstated, or
withheld-source claim.
- Supported claims: mention only material claims that survived re-checking.
- Evidence: use the footer labels from `${CLAUDE_PLUGIN_ROOT}/references/evidence-discipline.md`
when the audit depends on business/provider/customer or readiness evidence.
- Confidence: High, Medium, or Low, with the body matching the confidence.
Do not defend the prior answer. If the earlier answer was wrong, say what was
wrong and what the current evidence supports.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · warwick-bit/llm-accuracy | 100 | MIT |
| vinhnxv/rune | 95 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 36012633bce9 (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
- 90qstack-explain-for
Przepisuje poprzednią odpowiedź asystenta dla wskazanego czytelnika, takiego jak product manager, CEO czy CMO, opcjonalnie w ramach wskazanego produktu, gdy odpowiedź była zbyt techniczna lub zbyt wąska. - 90answer-audit
Audytuje odpowiedź systemu RAG wobec źródeł, zakładając błąd, dopóki każde twierdzenie nie zostanie potwierdzone dosłownie. - 84explain
Wyjaśnia prostym językiem pojęcie, kod, błąd, architekturę lub poprzednią odpowiedź asystenta, zaczynając od konkretnej analogii i bez żargonu. - 90qstack-ask-as-questions
Zamienia otwarte pytania z poprzedniej odpowiedzi asystenta w pytania strukturalne zadane przez narzędzie hosta, przeformułowane dla wskazanego czytelnika, z wersją ELI10, opcjami literowymi i jedną rekomendacją. - 90assistant-kpis-cost-per-outcome
Określa wskaźniki asystenta lub agenta AI: jakość w obszarach, przyjęcie i koszt poprawnego wyniku. - 100pa-onboarding
Prowadzi krok po kroku przez konfigurację nowego osobistego asystenta AI w OpenClaw, od założenia konta po pierwszą odpowiedź.