원클릭으로 Manus에서 모든 스킬 실행

doc-grounding

스타1

포크0

업데이트2026년 7월 8일 21:04

Docs-health grounding scan — doc claims verified against their SOURCE OF TRUTH: the code (README-vs-code, docstring params, examples that must run), the data (figures, tables, cited records), an original text (a rendering such as a translation, summary, or quote vs what it renders), or reality (versions, dates, external facts — verified REAL-TIME against MULTIPLE authoritative sources). Catches WRONG docs: fabricated claims, stale assertions, numbers that do not recompute. Triggers on: "/doc-grounding", "doc-grounding", "does the doc match the code", "verify doc claims", "fact-check this doc". Semantic (consent-gated Full) plus a cheap mechanical recompute layer (stated arithmetic, unit sanity). Offline or unverifiable degrades safe to "⚠️ unverified" — never a guess. Severity judged by context, never mechanical; claims needing formal verification escalate to CoalBoard.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

TheColliery

TheColliery/CoalLedger

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Doc-Grounding

Answer in the USER'S language; keep technical terms, commands, paths, and check ids verbatim.

Verify that a doc's claims match their source of truth. Report CONFIRMED mismatches only; never assert what you could not verify.

Parameters

SCOPE: named files (default when given) | touched doc files this session | whole repo docs (confirm first if > 20 files).
TIER: Quick = mechanical recompute only (~free, deterministic) · Full = semantic claim-by-claim verification (paid). Default from .coalledger.json quickVsFull; Full is always a separate consent.

Source of truth (general — pick per claim, never assume one kind)

claim about	source of truth
behavior / API / config	the code (read it; run an example only when cheap, safe, and consented)
figures / tables / results	the data or record it cites (compare verbatim)
a rendering of another text (translation, summary, quote)	the original text
external facts (versions, dates, prices, identifiers)	live authoritative sources — REAL-TIME and MULTI-source (cross-check several, never trust one), language-aware (fetch a source in the claim's language, or translate the claim to check it)

Method

Extract checkable claims (assertions a source can confirm or refute — skip opinions and intent).
Verify each against its source per the table. Mechanical layer first: RECOMPUTE stated arithmetic and unit/dimension sanity — deterministic, catches "2+2=5" for free.
Degrade safe: offline, source unreachable, or low-confidence language → mark ⚠️ unverified: check [source]. NEVER report an unverified claim as CONFIRMED; never fill the gap from memory.
Severity by CONTEXT (never a fixed map), then honor severityFloor: a wrong security or install instruction = CRITICAL; doc ≠ source on a surface readers actively rely on = HIGH; a stale minor claim = MEDIUM; cosmetic = LOW. publicMode: true raises the stakes of public-facing docs.

Escalation boundary (health ≠ correctness)

This canary verifies only what it can FETCH or RECOMPUTE. A claim needing formal verification — a proof, high-precision math, any error-not-allowed decision — is flagged and ESCALATED to CoalBoard (/coalboard), never adjudicated here.

Output

CONFIRMED table only; ⚠️ unverified and SUSPECTED go to separate lists, never the main table.

Fix mode (choice-gated)

After any report in an interactive session you MUST present this menu via your question tool (skip only when findings are zero or no user is present). NEVER auto-fix a live doc.

Apply safe fixes: only corrections the source states unambiguously (a version string, a figure copied wrong). Each fix: checkpoint (git stash/commit in a git repo; else copy the file aside — never assume git exists) -> apply -> re-read the changed lines -> revert on doubt.
Let me pick: list findings; the user selects.
Report only: exit unchanged.

NEVER auto-fix: any case where the DOC could be right and the source stale, any rewording of meaning, anything the source states ambiguously — offer options instead.

Multilingual

The recompute layer is language-agnostic. Semantic verification works in the doc's own language; a language handled poorly degrades to low-confidence ⚠️ unverified flags, never false alarms. A claim in one language with its source in another is verified cross-language (translate the claim, or fetch a language-matched source).

Problem report

If this canary misbehaves, OFFER to file it at https://github.com/TheColliery/CoalLedger/issues with a user-reviewed summary — never auto-submit.

이 저장소의 다른 Skills

같은 저장소

doc-structure

TheColliery/CoalLedger

Docs-health structure scan — broken internal links/anchors (GitHub-slug resolution, Thai/CJK safe), dead relative file links, heading hierarchy (skipped levels, multiple H1), GFM table shape (silently-dropped cells), orphan/undefined reference definitions, bare URLs in prose. Triggers on: "/doc-structure", "doc-structure", "broken links", "check docs structure", "doc health". Mechanical + deterministic: detection runs through the shipped CommonMark+GFM AST engine (never regex over raw markdown), so things that render fine are not flagged. Reports; fixes on request via choice-gated menu. Severity is judged by context, never mechanical.

2026-07-091

doc-consistency

TheColliery/CoalLedger

Docs-health contradiction scan — docs that disagree with EACH OTHER: the same fact stated two ways in two places, terminology drift (one thing under several names, or one name for several things), copy-paste blocks that were updated in one home and not the other, and cross-LANGUAGE drift (a doc and its counterpart in another language no longer say the same thing). Catches CONTRADICTORY doc sets. Triggers on: "/doc-consistency", "doc-consistency", "docs contradict", "terminology drift", "are the translations in sync". Semantic (paid, consent-gated — comparing meaning across files is judgment, not pattern-matching). Severity judged by context, never mechanical; CONFIRMED contradictions and SUSPECTED tensions reported separately.

2026-07-081

doc-leak

TheColliery/CoalLedger

Docs-health audience-safety scan — PROSE-level sensitive content in a doc meant to be public: internal URLs/hostnames/paths, personal data (names, emails, phone numbers in examples or logs), unpublished plans or roadmaps, private figures (pricing, customers), anything whose audience boundary the doc crosses. Catches LEAKED docs. Explicitly NOT a secret scanner: token-shaped secrets (API keys, passwords, private keys) are gitleaks' job — this canary hunts the sensitive PROSE that token scanners and gitignore both miss. Triggers on: "/doc-leak", "doc-leak", "is this safe to publish", "leak check", "anything private in this doc". Semantic, conservative by design: EVERY finding is SUSPECTED — public-vs-private is the human's judgment, never the tool's. Config-gated via the .coalledger.json docLeak key (default on; a private-only project turns it off).

2026-07-081

doc-quality

TheColliery/CoalLedger

Docs-health readability scan — two axes: QUALITY (bloat: filler, repetition, walls of text, buried leads; clarity: unexplained jargon, ambiguity, sentences that fight the reader) and language MECHANICS (typo, grammar, spelling, orthography — including script-level defects like decomposed characters that render right but break search, doubled spaces, wrong ellipsis/quote characters, spacing rules of the doc's language). Catches UNREADABLE or MALFORMED-LANGUAGE docs. Triggers on: "/doc-quality", "doc-quality", "tighten this doc", "proofread", "typos and grammar", "is this readable". Mechanical layer = deterministic mechanics (Unicode normalization, spacing, punctuation shape — ~free, report-only); semantic layer = bloat/clarity judgment + grammar in context (paid, consent-gated). Honors the project's own style/language rules where defined. Severity judged by context, never mechanical.

2026-07-081

doc-rot

TheColliery/CoalLedger

Docs-health staleness scan — content that time has made false: version strings older than the shipped version, dates and "last updated" stamps long past, stale badges, dead TODO/FIXME/"coming soon" markers, superseded instructions (steps for a surface that has since changed or gone), revalidate-by dates that have passed. Catches STALE docs — was true once, is not now. Triggers on: "/doc-rot", "doc-rot", "stale docs", "outdated docs", "is this doc still current". Mechanical layer = deterministic age-markers (versions, dates, TODO markers, badges — ~free, report-only); semantic layer = judging whether a flagged marker is actually rotten and spotting undated superseded content (paid, consent-gated). Severity judged by context, never mechanical; CONFIRMED and SUSPECTED reported separately.

2026-07-081

doc-standard

TheColliery/CoalLedger

Docs-health completeness scan — a doc measured against its KIND's standard: required sections present, public surface documented (every command/config key/exported API the code ships appears in the doc), the shape the genre expects (a README's install path, a security policy's reporting channel, a formal letter's parts). Catches INCOMPLETE docs. The standard is the project's OWN declared one first (a style guide, template, or pattern doc in the repo), else the kind's widely-accepted standard verified REAL-TIME against MULTIPLE authoritative sources. Triggers on: "/doc-standard", "doc-standard", "is this doc complete", "missing sections", "audit doc completeness". Mechanical layer (sections/hierarchy present via the AST — language-neutral) + semantic layer (completeness judgment, consent-gated). Offline degrades safe to "⚠️ unverified". Severity judged by context, never mechanical.

2026-07-081

name

doc-grounding

description

Doc-Grounding

Answer in the USER'S language; keep technical terms, commands, paths, and check ids verbatim.

Verify that a doc's claims match their source of truth. Report CONFIRMED mismatches only; never assert what you could not verify.

Parameters

SCOPE: named files (default when given) | touched doc files this session | whole repo docs (confirm first if > 20 files).
TIER: Quick = mechanical recompute only (~free, deterministic) · Full = semantic claim-by-claim verification (paid). Default from .coalledger.json quickVsFull; Full is always a separate consent.

Source of truth (general — pick per claim, never assume one kind)

claim about	source of truth
behavior / API / config	the code (read it; run an example only when cheap, safe, and consented)
figures / tables / results	the data or record it cites (compare verbatim)
a rendering of another text (translation, summary, quote)	the original text
external facts (versions, dates, prices, identifiers)	live authoritative sources — REAL-TIME and MULTI-source (cross-check several, never trust one), language-aware (fetch a source in the claim's language, or translate the claim to check it)

Method

Extract checkable claims (assertions a source can confirm or refute — skip opinions and intent).
Verify each against its source per the table. Mechanical layer first: RECOMPUTE stated arithmetic and unit/dimension sanity — deterministic, catches "2+2=5" for free.
Degrade safe: offline, source unreachable, or low-confidence language → mark ⚠️ unverified: check [source]. NEVER report an unverified claim as CONFIRMED; never fill the gap from memory.
Severity by CONTEXT (never a fixed map), then honor severityFloor: a wrong security or install instruction = CRITICAL; doc ≠ source on a surface readers actively rely on = HIGH; a stale minor claim = MEDIUM; cosmetic = LOW. publicMode: true raises the stakes of public-facing docs.

Escalation boundary (health ≠ correctness)

Output

CONFIRMED table only; ⚠️ unverified and SUSPECTED go to separate lists, never the main table.

Fix mode (choice-gated)

After any report in an interactive session you MUST present this menu via your question tool (skip only when findings are zero or no user is present). NEVER auto-fix a live doc.

Apply safe fixes: only corrections the source states unambiguously (a version string, a figure copied wrong). Each fix: checkpoint (git stash/commit in a git repo; else copy the file aside — never assume git exists) -> apply -> re-read the changed lines -> revert on doubt.
Let me pick: list findings; the user selects.
Report only: exit unchanged.

NEVER auto-fix: any case where the DOC could be right and the source stale, any rewording of meaning, anything the source states ambiguously — offer options instead.

Multilingual

Problem report

If this canary misbehaves, OFFER to file it at https://github.com/TheColliery/CoalLedger/issues with a user-reviewed summary — never auto-submit.