菜单
Use when reviewing agent output for factual accuracy -- detects fabricated URLs, unsourced statistics, fake citations, impossible future claims, and confident assertions without evidence
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use after completing a major work batch (plan execution, feature build, multi-task subagent run) to refresh code-graph-derived data in MEMORY.md/CONTEXT.md/SKILL_MEMORY, append dated narrative entries to TODO/CHANGELOG/lessons-learned, mine cross-project structural habits, and ratchet reusable permissions to global settings. Keeps project DNS in sync with reality.
Use when orchestrating multiple ghengis-skills across a single task -- coordinates pre-action validation, execution, post-action verification, and lesson extraction via a shared JSON scratchpad. TRIGGER when dispatching a subagent, finishing a major task, or when the user mentions "run the full chain", "validate this", "fire the pipeline", or explicitly invokes a chain by name.
Use to build and maintain an evolving understanding of user preferences, communication style, and working patterns — learns from interactions and adapts behavior over time
Use when starting a new project or organizing an existing one — creates self-documenting 4-layer project structure with MEMORY.md, CONTEXT.md, modular CLAUDE.md, and workspace routing
Use to accumulate and retrieve domain knowledge from past tasks — builds a grepable plain-text knowledge base that grows over time, enabling agents to learn from experience without vector databases
Use when reorganizing an existing project's file tree based on what actually connects to what — analyzes import graph, classifies layers, proposes an ideal target tree, validates the plan adversarially, then executes file moves with atomic per-language import rewrites. Always plan-first; never moves a file before the user confirms. TRIGGER on "reorganize this project", "clean up the file tree", "the structure is a mess", "make this match a real project layout", or after project-scaffold detects layout drift.
| name | hallucination-detector |
| description | Use when reviewing agent output for factual accuracy -- detects fabricated URLs, unsourced statistics, fake citations, impossible future claims, and confident assertions without evidence |
Signal-based detection of fabricated content in LLM output. Catches the four most common hallucination types: fabricated URLs, unsourced statistics, fake citations, and impossible future claims. Zero LLM cost, sub-millisecond latency -- pure regex pattern matching.
LLMs generate plausible-looking URLs from training data patterns. These URLs often look real but point to nonexistent pages, expired content, or entirely made-up domains.
Detection pattern:
Any URL (https?://...) in agent-generated text
Key distinction: URLs returned by tools (search results, API responses, browser navigation) are legitimate. URLs that the agent wrote from its own knowledge are suspect.
Common fabrication patterns:
https://docs.{product}.com/{plausible-path} -- documentation links that look right but 404https://github.com/{user}/{repo} -- repositories that do not existhttps://arxiv.org/abs/{number} -- paper IDs that are fabricatedhttps://{company}.com/blog/{date}/{slug} -- blog posts that were never writtenhttps://www.{news-site}.com/{year}/{month}/{headline} -- news articles that do not existRecovery:
LLMs generate precise-looking numbers to sound authoritative. Decimal percentages without a citation are a strong hallucination signal.
Detection pattern:
Any decimal percentage: \b(\d{1,3}\.\d+)\s*%
Examples: 73.4%, 2.7%, 94.2%, 156.8%
What this catches:
What this does NOT catch (intentionally):
Recovery:
LLMs commonly fabricate academic citations, including author names, publication years, and paper titles. These look highly credible but often reference papers that do not exist.
Detection patterns:
Pattern A -- Attribution phrases:
"according to [Author]"
"as shown/demonstrated/reported/found by [Author]"
"as shown by Dr. [Author] et al."
"as reported by [Author] and colleagues"
"as found by [Author] (2024)"
Pattern B -- Inline citations:
[Author] et al. (2024)
[Author] et al. [2024]
Pattern C -- Study references:
"a 2024 study by [Author]"
"a 2023 study by [Author]"
Where [Author] is a capitalized name (e.g., "Smith", "Johnson", "Zhang").
What this catches:
Recovery:
LLMs cannot predict the future but sometimes make definitive claims about upcoming events, release dates, or outcomes.
Detection patterns:
Definitive future language:
"will definitely..."
"will certainly..."
"will absolutely..."
"will guaranteed..."
"is guaranteed to..."
Specific future dates with product claims:
"will be released on/in/by [Month] [Year]"
"will launch on/in/by [Month] [Year]"
"will ship on/in/by [Month] [Year]"
What this catches:
What this does NOT catch (intentionally):
Recovery:
All four signal types default to "warn" severity. Escalate to "block" in these contexts:
| Context | Severity | Reason |
|---|---|---|
| Medical or health advice with fabricated statistics | Block | Could cause real harm |
| Legal citations that are fabricated | Block | Could lead to legal consequences |
| Financial data with made-up numbers | Block | Could lead to financial loss |
| URLs in output intended for end users | Block | Users will click broken links |
| Academic citations in research deliverables | Block | Destroys credibility |
| Casual conversation or brainstorming | Warn | Lower stakes, hallucinations are less harmful |
| Internal planning or reasoning | Warn | Not user-facing |
These request patterns are most likely to produce hallucinated content. Apply extra scrutiny to output generated from these prompts:
| Trigger Request | Why It Causes Hallucination |
|---|---|
| "Give me the exact statistics on..." | LLMs generate precise numbers to satisfy the "exact" requirement |
| "Find me 10 sources for..." | LLMs fabricate sources to fill the requested count |
| "What's the current data on..." | LLMs confuse training data with current data |
| "List the top 10 tools for..." | LLMs invent entries to complete the list |
| "According to research, what..." | LLMs fabricate citations to sound authoritative |
| "What will happen when..." | LLMs make definitive future claims to sound confident |
| "Give me a comprehensive list of..." | LLMs pad lists with fabricated entries to appear thorough |
| "What's the URL for..." | LLMs construct plausible URLs from patterns rather than admitting uncertainty |
| "Cite your sources" | LLMs fabricate citations retroactively to justify their output |
Mitigation for trigger requests: When you recognize a trigger pattern in the original instruction, proactively use tools (search, browse) to ground your output in verifiable data. Prefer "I don't have current data on this -- let me search" over fabricating an answer.
no_hallucination which provides the prompt-level instruction "Never fabricate URLs, citations, or statistics." Hallucination detector is the enforcement layer -- it checks whether the output actually followed that instruction.The most important principle: tool results are grounded, agent-generated text is not.
When in doubt about whether content came from a tool or from the agent's own generation, treat it as agent-generated and flag it.
$0. Zero LLM calls. All checks are regex pattern matching against the output text. Apply to every agent response, every subagent result, and every output before it reaches the user.
This skill participates in skill-chain-supervisor chains via the shared scratchpad at <project>/.claude/ghengis-chain/context.json.
Role in chain: Output scanner. Runs after any content-generating skill or agent.
Scratchpad subkey (namespaced writes): hallucination_detector.*
Reads (input scratchpad keys):
execution.resultWrites (output scratchpad keys):
hallucination_detector.fabricated_urls � list of URLs that look fake or uncheckablehallucination_detector.unsourced_stats � list of specific claims with no citationhallucination_detector.suspect_claims � list of confident assertions without evidenceSuccess criteria: all three lists empty OR claims verified against sources
When invoked as part of a chain, this skill MUST:
hallucination_detector.* namespace only � never overwrite another skill's subkeyhallucination_detector.error) rather than raisingWhen invoked standalone (not in a chain), scratchpad writes are optional but recommended for auditability.