Use deterministic rules for clear cases, classifiers for ambiguity
Apply deterministic rules to straightforward situations and auto-mode classifiers for ambiguous agent decision-making.
From Latent Space
Rules
One do or don't with the reason. 54 on file.
Coding18Writing14Agents and automation13Learning and teaching13Personal and everyday12Analysis and data10Planning and PM10Editing and humanizing7ChatGPT23Claude13Copilot9Gemini8Grok5Claude Code5
More rules
Just added
- Do not assume uninstalling Atlas deletes dataUninstalling Atlas does not remove all associated data. · OpenAI Help Center
- Verify AI-Generated Biological Designs with WatermarksEmbed watermarks in AI-generated biological designs to ensure authenticity and prevent misuse. · Google DeepMind Blog
- AI agents must not ignore limits to achieve goalsAI agents should respect defined operational boundaries and not bypass them to complete tasks. · BBC Technology
- Do not ask Claude to repeat its reasoning verbatimAvoid prompts that explicitly ask Claude to output its internal reasoning or chain of thought. · Claude Help Center
All 54
54 entries
Limit initial Grok Bot tests to fictional messages and your own address
Start Grok Bot testing with fictional messages and your own email to avoid unintended external communication.
From The Rundown AI
Set Grok Bot to ask before sending or forwarding email
Configure Grok Bot to request confirmation before sending or forwarding any email to prevent errors.
From The Rundown AI
Do not assume uninstalling Atlas deletes data
Uninstalling Atlas does not remove all associated data.
From OpenAI Help Center
Keep read operations synchronous for AI decision-making
Synchronize AI reads when its next decision depends on the returned data.
From Google Cloud Tech on YouTube
Verify AI-Generated Biological Designs with Watermarks
Embed watermarks in AI-generated biological designs to ensure authenticity and prevent misuse.
From Google DeepMind Blog
AI agents must not ignore limits to achieve goals
AI agents should respect defined operational boundaries and not bypass them to complete tasks.
From BBC Technology
Do not ask Claude to repeat its reasoning verbatim
Avoid prompts that explicitly ask Claude to output its internal reasoning or chain of thought.
From Claude Help Center
Keep every tool instant and put policy in one callback
Centralize policy checks in a single callback for AI agent tool execution.
From Google Cloud Tech on YouTube
Keep Tool Definitions Stable for GPT-6 Cache Reuse
Maintain stable tool definitions, schemas, and ordering to preserve reusable context in GPT-6.
From OpenAI News
Delete shared links to prevent future access
Remove shared links for ChatGPT conversations or tasks when no longer needed to stop future access.
From OpenAI Help Center
Apply higher standards to AI-generated code than human-written code.
AI code needs more rigorous checks than human code to ensure quality and maintainability.
From Simon Willison
Allow WebSocket traffic for full ChatGPT functionality
Permit WebSocket traffic over TCP port 443 to ensure ChatGPT features function correctly.
From OpenAI Help Center
Allowlist *.oaiusercontent.com to resolve specific ChatGPT errors
Allowlist *.oaiusercontent.com at the VPN or network security level to fix certain ChatGPT errors.
From OpenAI Help Center
Do not use event-triggered tasks for PHI in healthcare workspaces
Healthcare users must avoid transmitting or processing protected health information with event-triggered tasks.
From OpenAI Help Center
Use correct ChatGPT account for course progress and badges
Associate course progress and badges with the right ChatGPT account from the start.
From OpenAI Help Center
Do not include sensitive personal information in shared ChatGPT content
Avoid pasting health, financial, password, or account details into content that will be shared via ChatGPT links.
From OpenAI Help Center
Do Not Use ChatGPT for Medical Image Analysis
Avoid using ChatGPT to interpret specialized medical images or for medical advice.
From OpenAI Help Center
Manage Google Drive files in Google Drive
Control and delete Google Drive files directly from Google Drive, not ChatGPT.
From OpenAI Help Center
Obtain consent before recording others with ChatGPT
Always get permission before recording conversations involving other people.
From OpenAI Help Center
Specify chart types and columns for ChatGPT analysis
Guide ChatGPT's analysis by explicitly stating desired chart types and data columns.
From OpenAI Help Center
Check important information from ChatGPT
Always verify critical information provided by ChatGPT, as it can make mistakes.
From OpenAI Help Center
Complete age verification within 60 days in Italy
Users in Italy must verify age within 60 days to avoid feature limitations.
From OpenAI Help Center
Preserve experimental notification names for backward compatibility
Do not rename experimental notifications if consumers rely on them.
From OpenAI Developers Blog
Review and patch vulnerabilities in Claude Code
Use Claude Code to review proposed fixes for security vulnerabilities found by Claude Security.
From Claude Help Center
Use Opus for complex visualization tasks
For complex custom visuals, choose the Opus model for best performance.
From Claude Help Center
Include essential details when reporting issues for others
When reporting an issue on behalf of someone else, include their email, error messages, and start time.
From Claude Help Center
Keep one conversation per support issue
Consolidate all communication for a single problem into one support conversation.
From Claude Help Center
Limit Claude Accounts to Three Per Phone Number
A single phone number can verify a maximum of three Claude accounts.
From Claude Help Center
Use a Long-Term Accessible Phone Number for Claude Verification
Verify your Claude account with a phone number you will have long-term access to.
From Claude Help Center
AI use should prioritize human control and oversight
Maintain human control over AI systems, ensuring people are in charge of their day-to-day activities.
From Microsoft Source
Copilot approval does not count by default
Copilot's approval assessment alone does not satisfy merge requirements unless explicitly enabled.
From GitHub Changelog: Copilot
Copilot approval is dismissed on new commits
If new commits are pushed after Copilot approves, its approval is automatically dismissed.
From GitHub Changelog: Copilot
Multipart form data: file parameter must be last
When using multipart/form-data, ensure the 'file' parameter is always placed after all other parameters.
From xAI Docs: Guides
Pause AI activity on critical security boundary violation alerts
Immediately pause AI activity if a critical security boundary violation cannot be disproven within 30 minutes.
From OpenAI News
Do not use presencePenalty, frequencyPenalty, stop with reasoning models
Avoid using presencePenalty, frequencyPenalty, and stop parameters with Grok reasoning models to prevent errors.
From xAI Docs: Guides
Custom voices are team-scoped and private
Your custom voices are only available to your team and not shared with other users.
From xAI Docs: Guides
Prioritize understanding over speed with AI coding tools
Focus on learning concepts when using AI coding assistants, not just task completion speed.
From Anthropic Research
Recognize and decline fraudulent content creation
AI models must recognize and decline attempts to create fraudulent content.
From Anthropic News
Recognize and decline malicious code generation
AI models must recognize and decline attempts to generate malicious code.
From Anthropic News
Block arbitrary code execution tools in Claude Code auto mode
Disable broad code execution tools when entering auto mode to ensure the classifier sees potentially dangerous commands.
From Anthropic Engineering
Agents can create new work items
Allow AI agents to file new issues when they identify improvements or related tasks outside the current scope.
From OpenAI News
Maintain human oversight for critical AI-assisted tasks
Always double-check AI outputs for accuracy and relevance before using them for important work.
From OpenAI News
Verify AI-generated advice for critical decisions
Always consult qualified professionals for legal, medical, or financial advice, not just AI.
From OpenAI News
Use AI for system-level context in code reviews
Leverage AI tools that provide system-wide context to enhance code review beyond static analysis.
From OpenAI News
Default Claude Code to read-only access
Claude Code operates on a permission-based model, defaulting to read-only access.
From Anthropic Engineering
Prioritize text-based response evaluation for LLM bias
Focus LLM bias evaluation on text outputs, as they represent the majority of everyday usage.
From OpenAI News
Reason about when and how to use tools for better AI performance
AI models should be trained to reason about tool usage, not just execute tools.
From OpenAI News
Avoid judgmental language and excessive verbosity in model responses
Model responses should be concise and avoid judgmental or overly verbose language.
From OpenAI News
Model refusal should contain a brief apology and state inability to comply
When a model refuses a request, it should include a short apology and clearly state it cannot fulfill the request.
From OpenAI News
Model responses to sensitive topics should be empathetic but decline to comply
For sensitive topics, models should offer an empathetic apology before declining the request.
From OpenAI News
Ask clarifying questions when necessary in interactive settings
Seek more information from users to ensure accurate and relevant responses.
From OpenAI News
Be helpful without overstepping on regulated topics
Provide information on regulated topics but do not give professional advice.
From OpenAI News
Use natural language prompts for GPT-3 programming
Program GPT-3 by providing natural language prompts and examples, not traditional code.
From OpenAI News
Nothing matches. Clear a filter or try a shorter word.