anti-hallucination
CRITICAL SAFETY SKILL — Verify ALL technical claims, API signatures, library methods, code behavior, and factual statements…
Conduct rigorous research with proper citations (DOI, arXiv, PMID) and source triangulation. MUST BE USED when user asks: "what is SOTA", "recent developments", "compare X vs Y", "is it true that", "research says", "latest papers on", "scientific evidence", "studies show",
$ npx -y skills add Aedelon/claude-code-blueprint --skill research-protocol --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/research-protocolContext preview
The summary Claude sees to decide when to auto-load this skill.
Conduct rigorous research with proper citations (DOI, arXiv, PMID) and source triangulation. MUST BE USED when user asks: "what is SOTA", "recent developments", "compare X vs Y", "is it true that", "research says", "latest papers on", "scientific evidence", "studies show",
name: research-protocol description: |- Conduct rigorous research with proper citations (DOI, arXiv, PMID) and source triangulation. MUST BE USED when user asks: "what is SOTA", "recent developments", "compare X vs Y", "is it true that", "research says", "latest papers on", "scientific evidence", "studies show", "state of the art", "literature review", "find papers", "academic research", "benchmark results", "who published", "when was X released", "current best", "what does the research say", "evidence for", "peer reviewed". Searches multiple sources, evaluates reliability, states confidence level. NOT for verifying API signatures (use anti-hallucination) or general web search (use WebSearch directly). allowed-tools: - WebSearch - WebFetch - mcp__fetch__fetch - mcp__huggingface__get-paper-info - mcp__huggingface__get-daily-papers - mcp__huggingface__search-models - Read
1. SCOPE → Define question precisely 2. SEARCH → Multiple sources, cross-reference 3. EVALUATE → Assess source reliability 4. SYNTHESIZE → Combine findings coherently 5. CITE → Provide proper attribution 6. ASSESS → State confidence level
| Priority | Source Type | Reliability | Citation Format | |----------|-------------|-------------|-----------------| | 1 | Peer-reviewed papers | Highest | [Author et al., Year] arXiv:XXXX | | 2 | Official documentation | High | Docs [Library] vX.X | | 3 | Conference proceedings | High | [Conf Year] Paper Title | | 4 | Established tech blogs | Medium-High | [Org] Blog (Date) | | 5 | GitHub repos with citations | Medium | GitHub [repo] | | 6 | Stack Overflow (verified) | Medium | SO [answer-id] | | 7 | General web content | Low | Mention skepticism |
1. mcp__huggingface__get-daily-papers() → Recent papers
2. WebSearch("[topic] SOTA 2024")
3. WebSearch("[topic] benchmark comparison")
4. mcp__huggingface__search-models(tags: "[task]")1. WebSearch("paper title arxiv")
2. mcp__huggingface__get-paper-info(arxiv_id: "XXXX.XXXXX")1. mcp__context7__resolve-library-id() first 2. Then research protocol for context
[Author et al., Year] "Title" - arXiv:XXXX.XXXXX [Author et al., Year] "Title" - DOI:10.XXXX/XXXXX [Author et al., Year] "Title" - PMID:XXXXXXXX
According to [Library] v[X.X] documentation: ... According to [Framework] docs (2024): ...
According to [Source Name] (Date): ... Per [Organization] blog (Month Year): ...
| Confidence | Criteria | Response Style | |------------|----------|----------------| | **HIGH** | 3+ concordant sources, peer-reviewed | State as established fact | | **MEDIUM** | 1-2 reliable sources | Add "according to [source]" caveat | | **LOW** | Conflicting sources | Present multiple views | | **UNKNOWN** | No reliable sources | "I don't know" |
When sources disagree:
## Divergent Viewpoints **Position A** (Source 1, Source 2): [Description] **Position B** (Source 3): [Description] **Assessment**: [Which seems more credible and why]
## Summary [Main findings - 3-5 lines max] ## Details [Expanded information with inline citations] ## Sources 1. [Citation 1 with link/DOI] 2. [Citation 2 with link/DOI] 3. ... ## Confidence Level: [HIGH/MEDIUM/LOW] [1-2 sentence justification] ## Limitations - [What couldn't be verified] - [Potential biases in sources] - [Recency concerns if applicable]
Automatically engage this protocol when user asks about:
A production-grade Claude Code configuration: skills, agents, hooks, rules, and permissions working as a system, not as isolated features. Companion repo for the article: "I Turned Claude Code Into an Operating System. Here's the Blueprint."
CRITICAL SAFETY SKILL — Verify ALL technical claims, API signatures, library methods, code behavior, and factual statements…
Multi-agent brainstorming and strategic thinking on any complex topic. Explores the subject deeply, then spawns specialized…
Reference patterns for REST APIs, pytest/vitest testing, Docker multi-stage builds, GitHub Actions CI/CD, PostgreSQL, TypeScript…
Ship workflow: review changes, generate conventional commit messages, push, and create PRs. MUST BE USED when user says:…
Debug errors systematically by searching first, then analyzing, then proposing verified solutions. MUST BE USED when user…
Proactive security audit: OWASP top 10, dependency vulnerabilities, secrets detection, input validation, auth patterns, and…