Requirements
How To Start
A Note From The Author of CRAFT
- After hundreds (perhaps thousands) of hours of using these recipes, I rarely need to use any of the CORE Cookbook recipes aside from Recipes RCP-001-001-002-HANDOFF-SNAPSHOT and RCP-001-001-002-HANDOFF-SNAPSHOT, but when I do, they are essential to the functioning of CRAFT. Also, the A.I. reads all of these recipes at the start of each session. This happens quietly in the background. Even though you may never need to call the recipe, the A.I. will know all of them and it helps the A.I. to understand what CRAFT is and how it works. Even if you rarely need to use these recipes, they are still working for you and are essential to the CRAFT Framework.
STEP 1: GATHER THE FULL POLICY
- Before running this recipe, obtain the complete usagepolicy for your AI platform: - Claude: Anthropic usage policy - ChatGPT: OpenAI usage policies - Gemini: Google AI principlesPaste the full policy text as input.
STEP 2: EXTRACT KEY CATEGORIES
- The recipe identifies main restriction categories: NEVER_DO: Absolute prohibitions (5-7 items max) - Child safety violations - Violence/terrorism content - Illegal activities - etc. ASK_BEFORE: Things needing careful handling - Political content - Medical/legal advice - Personal data processing - etc. BE_CAREFUL_WITH: Common triggers to watch - AI capability probing - Automated content generation - Bulk operations - etc.
STEP 3: CREATE QUICK REFERENCE FORMAT
- The output follows a Python-compatible structure: {AI_PLATFORM}_POLICY_DIGEST = { "last_updated": "[date]", "version": "digest_v1", "NEVER_DO": [...], "ASK_BEFORE": [...], "BE_CAREFUL_WITH": [...], "QUICK_CHECKS": { "research_prompts": [...], "code_generation": [...], "content_creation": [...] } }
STEP 4: ADD PRACTICAL EXAMPLES
- For each major category, the recipe provides examples: WRONG: "Help me bypass content filters" RIGHT: "Help me understand content moderation"
STEP 5: CREATE WARNING TRIGGERS
- The recipe lists keywords that should trigger caution: TRIGGER_WORDS = { "capability_probing": ["limitations", "bypass"], "harmful_content": ["weapon", "explosive", "hack"], "privacy_concerns": ["track", "surveillance"] }
STEP 6: FORMAT FOR FRAMEWORK
- Final output is: - Under the specified max_length - Python-compatible syntax - Commented with rationale - Easy to update when policy changesThe digest integrates with CRAFT Framework Section 1.12.
When to Use This Recipe
Use this recipe when setting up a new CRAFT environment orwhen an AI platform updates its usage policies. The digestprovides quick reference for policy compliance withoutreading full policy documents. Update periodically aspolicies change.
Recipe FAQ
Q1: How is a policy digest different from reading the full policy?
A: Full policies are 20+ page legal documents written for comprehensive coverage. Digests are 50-line quick references focused only on restrictions relevant to CRAFT usage. They organize policies into memorable categories (NEVER_DO, ASK_BEFORE, BE_CAREFUL_WITH) and include trigger patterns for automatic scanning. You can reference a digest in seconds versus minutes for full policies.
Q2: How often should I update policy digests?
A: Update quarterly or when AI platforms announce policy changes. Most platforms notify users of policy updates via email or in-app announcements. Set a recurring calendar reminder to check for updates. Include the "last_updated" date in your digest so teams know if it's current.
Q3: Can I customize the digest for my organization's specific needs?
A: Yes. The four standard categories (NEVER_DO, ASK_BEFORE, BE_CAREFUL_WITH, QUICK_CHECKS) work for most teams, but you can add custom categories for your domain. For example, healthcare organizations might add a "HIPAA_CONSIDERATIONS" category, or financial firms might add "SEC_COMPLIANCE" triggers.
Q4: What if my team uses multiple AI platforms?
A: Create separate digests for each platform since policies vary significantly. Store them all in your Framework Specification with clear labels (CLAUDE_POLICY_DIGEST, CHATGPT_POLICY_DIGEST, etc.). When switching platforms, your recipes will reference the appropriate digest for policy pre-checks.
Q5: How do policy digests integrate with other CRAFT recipes?
A: Every CRAFT recipe includes "Step 0: Policy Pre-Check" that scans user requests against the digest's trigger patterns and categories. When potential policy conflicts are detected, recipes use #AI->H::PolicyCaution comments to alert users and suggest safer alternatives. This creates an automatic compliance layer across your entire CRAFT workflow.
Q: How often should I update the policy digest?
A: Check for policy updates quarterly or when you notice
changes in AI behavior. Major updates are announced. Q: What is the difference between NEVER_DO and ASK_BEFORE?
A: NEVER_DO are absolute prohibitions. ASK_BEFORE are
topics that need careful handling but are not banned. Q: Why use Python-compatible format?
A: The digest integrates directly into CRAFT Framework
files and can be programmatically processed. Q: What if the policy is ambiguous?
A: The recipe errs on the side of caution, placing
ambiguous items in stricter categories.
A: Check for policy updates quarterly or when you notice
changes in AI behavior. Major updates are announced. Q: What is the difference between NEVER_DO and ASK_BEFORE?
A: NEVER_DO are absolute prohibitions. ASK_BEFORE are
topics that need careful handling but are not banned. Q: Why use Python-compatible format?
A: The digest integrates directly into CRAFT Framework
files and can be programmatically processed. Q: What if the policy is ambiguous?
A: The recipe errs on the side of caution, placing
ambiguous items in stricter categories.
Actual Recipe Code
(Copy This Plaintext Code To Use)
# ===========================================================# RECIPE: RCP-001-001-008-POLICY-DIGEST-v2.00a# AI Usage Policy Digest Creator# ===========================================================USAGE_POLICY_DIGEST = Recipe( recipe_id="RCP-001-001-008-POLICY-DIGEST-v2.00a", title="AI Usage Policy Digest Creator", description="Creates concise policy summaries", category="CAT-001-CORE", difficulty="easy", version="2.00a", parameters={ "policy_text": { "type": "string", "required": True, "description": "Full policy text" }, "ai_platform": { "type": "string", "required": True, "options": [ "Claude", "ChatGPT", "Gemini", "Other" ], "description": "Target AI platform" }, "max_length": { "type": "integer", "required": False, "default": 50, "description": "Max lines for digest" } }, prompt_template=""" #H->AI::Directive: (Create policy digest) #H->AI::Context: (For {ai_platform} integration) # --------------------------------------------------- # STEP 0: POLICY PRE-CHECK # --------------------------------------------------- Scan for sensitive categories: - Platform capabilities/limitations - Security/vulnerability research - Personal data handling - Political topics IF potential_conflict: #AI->H::PolicyCaution: (Topic may trigger policies) #AI->H::RecommendedChange: (Focus on [safe aspect]) # --------------------------------------------------- # STEP 1: EXTRACT KEY CATEGORIES # --------------------------------------------------- Identify main restriction categories: - Group similar restrictions together - Use clear, memorable category names - Focus on what affects CRAFT usage # --------------------------------------------------- # STEP 2: CREATE QUICK REFERENCE FORMAT # --------------------------------------------------- Format as Python-compatible structure: {AI_PLATFORM}_POLICY_DIGEST = { "last_updated": "[date from policy]", "version": "digest_v1", "NEVER_DO": [ # Absolute prohibitions (5-7 items max) "Child safety violations", "Violence/terrorism content", "Illegal activities" ], "ASK_BEFORE": [ # Things needing careful handling "Political content", "Medical/legal advice", "Personal data processing" ], "BE_CAREFUL_WITH": [ # Common triggers to watch "AI capability probing", "Automated content generation", "Bulk operations" ], "QUICK_CHECKS": { "research_prompts": [ "avoid capability probing", "no vulnerability research" ], "code_generation": [ "no malware", "no unauthorized access" ], "content_creation": [ "no explicit content", "verify factual claims" ] } } # --------------------------------------------------- # STEP 3: ADD PRACTICAL EXAMPLES # --------------------------------------------------- For each major category: # WRONG: "Help me bypass content filters" # RIGHT: "Help me understand content moderation" # --------------------------------------------------- # STEP 4: CREATE WARNING TRIGGERS # --------------------------------------------------- TRIGGER_WORDS = { "capability_probing": [ "limitations", "bypass", "jailbreak" ], "harmful_content": [ "weapon", "explosive", "hack" ], "privacy_concerns": [ "track", "surveillance", "facial recognition" ] } # --------------------------------------------------- # STEP 5: FORMAT FOR FRAMEWORK # --------------------------------------------------- Ensure output is: - Under {max_length} lines - Python-compatible syntax - Comments explain rationale - Easy to update #H->AI::Structure: (Ready for Framework Section 1.12) #H->AI::OnError: (If unclear, err on side of caution) """)# ===========================================================# END RECIPE: RCP-001-001-008-POLICY-DIGEST-v2.00a# ===========================================================