菜单
Create a Model/System Card (MCARD) document through an interactive step-by-step flow. Guides the user through each section with specific questions and example responses.
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
Create an ADR (Architecture Decision Record) for significant architectural decisions. Requires human review.
Create an AIDEC (AI Decision) document to record a technical decision with alternatives considered.
Create an AILOG (AI Action Log) document for the current changes. Quick shortcut for the most common document type.
Execute an external audit of a Charter inside an auditor-side CLI (gemini-cli, claude-cli, copilot-cli, codex-cli). Reads the resolved audit prompt from the canonical location, audits with tool use, writes the report. Operator invokes one instance per CLI.
Generate the unified audit prompt for a Charter at the canonical filesystem location. The operator then opens N auditor-side CLIs (gemini-cli, claude-cli, copilot-cli, etc.) and invokes /straymark-audit-execute in each — no copy/paste. Counterpart of /straymark-audit-review.
Consolidate N external auditor reports into a critical review document with verdicts, remediation plan, and auditor ratings. Then merge the external_audit YAML block into the Charter telemetry. Counterpart of /straymark-audit-prompt and /straymark-audit-execute.
| name | straymark-mcard |
| description | Create a Model/System Card (MCARD) document through an interactive step-by-step flow. Guides the user through each section with specific questions and example responses. |
| allowed-tools | Read, Write, Glob, Bash(git status *, git log *, git diff *, date *, wc *) |
Create AI Model/System Card documentation through an interactive guided flow.
When invoked, follow these steps:
If the user specified a model name (e.g., /straymark-mcard GPT-4o), use it as the model name and proceed to step 2 asking only for the model type.
If no parameter is given, proceed to step 2 asking for both model name and type.
Ask the user:
╔══════════════════════════════════════════════════════════════════╗
║ StrayMark MCARD — Model/System Card ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ Step 1/7: Model Identity ║
║ ║
║ Please provide the following: ║
║ ║
║ 1. Model name: [e.g., "GPT-4o", "BERT-base-uncased"] ║
║ 2. Model type: ║
║ • LLM — Large Language Model ║
║ • classifier — Classification model ║
║ • regressor — Regression model ║
║ • generator — Generative model (image, audio, etc.) ║
║ • recommender — Recommendation system ║
║ • other — Specify ║
║ 3. Provider: [e.g., "OpenAI", "Google", "Hugging Face"] ║
║ 4. Version: [e.g., "2024-05-13", "v1.0", "gpt-4o-2024-05-13"] ║
║ 5. License: [e.g., "Proprietary", "Apache 2.0", "MIT"] ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Wait for user response before proceeding.
Ask the user:
╔══════════════════════════════════════════════════════════════════╗
║ Step 2/7: Intended Use ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ 1. Primary intended uses: ║
║ Example: "Code generation", "Customer support chatbot", ║
║ "Sentiment analysis on product reviews" ║
║ ║
║ 2. Primary intended users: ║
║ Example: "Internal engineering team", "End users via API", ║
║ "Data science team" ║
║ ║
║ 3. Out-of-scope uses (what the model should NOT be used for): ║
║ Example: "Medical diagnosis", "Legal advice", ║
║ "Autonomous decision-making without human review" ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Wait for user response before proceeding.
Ask the user:
╔══════════════════════════════════════════════════════════════════╗
║ Step 3/7: Training Data ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ 1. Dataset name: [e.g., "Common Crawl", "Internal corpus v3"] ║
║ 2. Source: [e.g., "Web scraping", "Licensed data provider"] ║
║ 3. Size: [e.g., "1.5T tokens", "500K samples", "200GB"] ║
║ 4. Collection methodology: ║
║ Example: "Web crawling with quality filters" ║
║ 5. Preprocessing steps: ║
║ Example: "Deduplication, PII removal, language filtering" ║
║ 6. Known limitations in the data: ║
║ Example: "English-centric, underrepresents African languages" ║
║ 7. PII assessment: ║
║ Example: "PII filtered using regex + NER; residual risk low" ║
║ 8. Data license: [e.g., "CC-BY-4.0", "Proprietary"] ║
║ ║
║ Type "unknown" or "N/A" for fields you cannot fill. ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Wait for user response before proceeding.
Ask the user:
╔══════════════════════════════════════════════════════════════════╗
║ Step 4/7: Performance & Evaluation ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ Performance Metrics (provide at least one): ║
║ Format: Metric | Value | Test Dataset | Conditions ║
║ Example: "Accuracy | 92.3% | MMLU | 5-shot" ║
║ "F1 | 0.87 | SQuAD v2 | zero-shot" ║
║ "BLEU | 34.2 | WMT-22 en-de | beam search k=5" ║
║ ║
║ Disaggregated evaluation (optional): ║
║ Format: Subgroup | Metric | Value ║
║ Example: "English | Accuracy | 95.1%" ║
║ "Spanish | Accuracy | 88.4%" ║
║ ║
║ Type "unknown" if metrics are not available. ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Wait for user response before proceeding.
Ask the user:
╔══════════════════════════════════════════════════════════════════╗
║ Step 5/7: Bias, Security & Ethics ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ Bias & Fairness (if evaluated): ║
║ Format: Group | Metric | Performance | Mitigation ║
║ Example: "Female | Toxicity | 0.02 | Content filter applied" ║
║ Type "not evaluated" if no bias analysis was performed. ║
║ ║
║ Security Concerns: ║
║ • Known vulnerabilities: [None / describe] ║
║ • Adversarial robustness: [Low / Medium / High] ║
║ • Prompt injection risk: [Low / Medium / High / N/A] ║
║ • Data poisoning risk: [Low / Medium / High] ║
║ • Model extraction risk: [Low / Medium / High] ║
║ ║
║ Ethical Considerations: ║
║ • Was sensitive data used in training? [Yes / No / Unknown] ║
║ • Human subjects involved? [Yes / No / Unknown] ║
║ • Dual-use potential? [Yes / No] ║
║ • Societal impact notes: [Brief description] ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Wait for user response before proceeding.
Ask the user:
╔══════════════════════════════════════════════════════════════════╗
║ Step 6/7: Environmental Impact ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ 1. Training energy (kWh): [e.g., "1,287 kWh", "Unknown"] ║
║ 2. CO2 equivalent (tons): [e.g., "0.58 tCO2", "Unknown"] ║
║ 3. Hardware used: [e.g., "8x NVIDIA A100 80GB"] ║
║ 4. Training duration: [e.g., "72 hours", "Unknown"] ║
║ 5. Inference cost: [e.g., "$0.005 per 1K tokens", "Unknown"] ║
║ 6. Region / Grid carbon intensity: [e.g., "US-East", "Unknown"] ║
║ ║
║ Type "unknown" for fields you cannot fill. ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Wait for user response before proceeding.
Ask the user:
╔══════════════════════════════════════════════════════════════════╗
║ Step 7/7: Limitations & Recommendations ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ Known limitations: ║
║ Example: "Poor performance on low-resource languages", ║
║ "Context window limited to 128K tokens", ║
║ "Hallucination rate ~5% on factual queries" ║
║ ║
║ Known failure modes: ║
║ Example: "Produces incorrect math on multi-step problems", ║
║ "May refuse safe queries due to over-filtering" ║
║ ║
║ Recommendations for deployers: ║
║ Example: "Implement output filtering for PII", ║
║ "Use human-in-the-loop for high-stakes decisions", ║
║ "Monitor for performance drift monthly" ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Wait for user response before proceeding.
Based on the gathered information, determine:
EU AI Act risk level: Classify based on intended use:
unacceptable: social scoring, real-time biometric identificationhigh: employment decisions, credit scoring, law enforcement, educationlimited: chatbots, emotion recognition, deepfake generationminimal: spam filters, game AI, search optimizationnot_applicable: no EU deployment plannedNIST GenAI risks: Select applicable risks from:
privacy, bias, confabulation, cbrn, dangerous_content, environmental, human_ai_config, information_integrity, information_security, intellectual_property, obscene_content, value_chain
Overall risk level: low, medium, high, or critical
Read .straymark/config.yml to determine language:
language: en # or es
Use template path based on language:
en (default): .straymark/templates/TEMPLATE-MCARD.mdes: .straymark/templates/i18n/es/TEMPLATE-MCARD.mdDetermine the next sequence number:
# Get current date
date +%Y-%m-%d
# Find existing MCARD documents for today
ls .straymark/09-ai-models/MCARD-$(date +%Y-%m-%d)-*.md 2>/dev/null | wc -l
ID format: MCARD-YYYY-MM-DD-NNN
Display a summary and ask for confirmation:
╔══════════════════════════════════════════════════════════════════╗
║ StrayMark MCARD — Summary ║
╠══════════════════════════════════════════════════════════════════╣
║ ║
║ Model: [model_name] ([model_type]) ║
║ Provider: [provider] — Version: [version] ║
║ EU AI Act Risk: [risk classification] ║
║ NIST GenAI Risks: [applicable risks] ║
║ Overall Risk: [risk level] ║
║ ║
║ Proposed filename: ║
║ MCARD-YYYY-MM-DD-NNN-[model-name-slug].md ║
║ ║
║ Location: ║
║ .straymark/09-ai-models/ ║
║ ║
║ Review required: YES ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
Confirm creation? [Y/n]:
Wait for user confirmation before proceeding.
YYYY-MM-DD -> Current dateNNN -> Sequence number (001, 002, etc.)[agent-name] -> claude-code-v1.0[Model Name] -> User-provided model namereview_required: true in the frontmatter.straymark/09-ai-models/MCARD-YYYY-MM-DD-NNN-[model-name-slug].mdAfter creation, display:
StrayMark MCARD created:
.straymark/09-ai-models/MCARD-YYYY-MM-DD-NNN-[description].md
Model: [model_name] ([model_type])
Provider: [provider]
Review required: yes
Risk level: [risk_level]
EU AI Act: [classification]
.straymark/09-ai-models/ directory: Create it before saving[To be determined][Information not publicly available] and note in limitations.straymark/config.yml: Default to English (en)Terminal compatibility: If the terminal does not support box-drawing characters (Unicode), use plain-text formatting with dashes and pipes instead (e.g.,
+--+instead of╔══╗).