Menü
Dual-dimension document quality gate — presentation rubric + hallucination grounding check
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Convert PDF/EPUB library to Markdown and generate Obsidian MOC notes
Hook-based compaction suggestions at logical task boundaries
Context window management — track spend, decide when to compact, preserve state
Session-start orientation — loads context, surfaces learnings, confirms registry
Quality and semantic review — catches what automated tools miss
Planner → Architect → Critic deliberation loop — produces a formally validated ADR
Basierend auf der SOC-Berufsklassifikation
| name | doc-review |
| description | Dual-dimension document quality gate — presentation rubric + hallucination grounding check |
| version | 0.1.0 |
| level | 2 |
| triggers | ["/doc-review","review this doc","check for hallucinations","hallucination check","doc quality check"] |
| context_files | ["context/user.md"] |
| steps | [{"name":"Scope","description":"Identify the document type and its stated sources. If no sources are declared, ask the user before proceeding."},{"name":"Presentation Pass","description":"Evaluate structure, clarity, and completeness against the rubric."},{"name":"Hallucination Pass","description":"For each factual claim, verify it against the stated source material. Flag ungrounded claims."},{"name":"Verdict","description":"PASS / REVISE / BLOCK with specific, line-referenced feedback."}] |
Two-pass quality gate for generated documents. Run before delivering any report, PRD, ADR, email, research summary, wiki page, or other prose output where accuracy matters.
Without explicit review, generated documents can pass a surface quality check while containing factual claims that were hallucinated — plausible-sounding assertions with no grounding in the source material. Claude Code averages 10+ hallucinations per generated research paper. A one-pass review catches presentation issues but misses grounding failures.
The two-pass structure separates these concerns: Presentation checks quality; Hallucination checks truth. Both must pass independently.
For long documents (PRDs, ADRs, research summaries), use the overview-anchor pattern before writing the full document:
This reduces hallucination risk: the overview forces explicit source declaration upfront, and each section of the full document is anchored to a confirmed claim.
Evaluate the document against this rubric:
| Dimension | Score (1-5) | Criteria |
|---|---|---|
| Structure | Clear sections, logical flow, no orphaned content | |
| Clarity | No ambiguous claims, jargon explained, terms consistent | |
| Completeness | All stated goals addressed, no missing sections | |
| Audience fit | Appropriate for the stated reader (technical vs. non-technical) |
Score 1-2 on any dimension = REVISE. Score 1 on any = BLOCK.
Common presentation failures:
For each factual claim in the document:
Claim categories:
| Category | Grounding requirement |
|---|---|
| Statistics / numbers | Must match source verbatim or cite approximate |
| Named entities (people, orgs, tools) | Must be verifiable from source or prior session context |
| Causal claims ("X causes Y") | Must be explicitly stated in source, not inferred |
| Recommendations | Must follow logically from verified claims |
| Quotes | Must be exact; partial quotes must not change meaning |
Ungroundable claims (not necessarily wrong, but unverifiable):
Flag ungroundable claims separately from hallucinations. They may be valid but must be surfaced for the user to validate.
PASS: Document meets presentation standards and all factual claims are grounded. State any minor observations that do not require changes.
REVISE:
Issue: [description]
Location: [section or line reference]
Dimension: [presentation / hallucination / ungrounded]
Fix: [specific recommendation]
Blocking: No
BLOCK:
Issue: [description]
Location: [section or line reference]
Dimension: [presentation / hallucination / ungrounded]
Impact: [what the reader would believe incorrectly]
Fix: [specific recommendation]
Blocking: YES
Hallucination findings that would mislead the reader are always BLOCK, not REVISE.
| Document type | Expected sources |
|---|---|
| Research summary | Source papers/URLs declared upfront |
| PRD | User-stated requirements from deep-interview or conversation |
| ADR | Decision drivers stated in the ADR's Context section |
| Email / client report | Facts from session context or user-provided data |
| Wiki page | Library source file (PDF/EPUB path) |
| Slide content | Source document or user-stated outline |
If the document type does not match a known pattern, ask the user to declare sources before running Pass 2. Do not guess at source material.
Do not combine doc-review with code-review. They are separate passes with different criteria.
Do not skip Pass 2 because "the document looks reasonable." Plausible hallucinations are the most dangerous kind.
Do not flag ungrounded claims as hallucinations. They are different: a hallucination contradicts the source; an ungrounded claim has no source to contradict.