This keeps it short, actionable, and hints at the “why” (protection) without overload. The “why” can live in a short intro section, explaining risks like data training, regret from oversharing, or escalating content, then dive into setup steps.
Guide Structure Overview
Divide into three layered sections: account-wide settings (biggest environment change), project/custom instructions (next layer), and per-prompt add-ons (quick wins). This creates a “defense in depth” where higher layers reinforce lower ones. Include your pasted examples as ready-to-copy prompts.
1. Account-Wide Settings
Most LLMs (ChatGPT, Claude, Google Gemini, Perplexity) let you set global instructions or security defaults that apply everywhere. These change the core environment first.
| LLM Platform | How to Set Global Guardrails | Key Settings to Enable |
|---|---|---|
| ChatGPT (OpenAI) | Go to Settings > Custom Instructions | Toggle “Privacy” mode; add system prompt: “Always warn on PII/oversharing. Block explicit content. Suggest rephrasing.” Enable data controls to opt out of training. |
| Claude (Anthropic) | Account Settings > Default Instructions | Set “Strict safety” mode; input: “Detect sensitive info and pause. No NSFW. Offer learning alternatives.” |
| Gemini (Google) | Profile > AI Preferences | Enable “Safe mode” and “Family Link” if applicable; add: “Flag personal details. Redirect risky topics educationally.” |
| Grok (xAI) | Settings > Behavior | Custom system: “Protect user privacy; warn on risks; no graphic content.” |
| General Tip | Check “Data & Privacy” tab | Opt out of model training; enable content filters; set age-appropriate defaults. |
Why it works: These apply automatically, reducing reliance on per-use thinking environment shifted upfront.
2. Project / Custom Instructions (Mid-Layer)
For Perplexity”projects,” Gemini “gems,” or all saved chats, layer in detailed rules. Copy-paste these templates.
- Core AI Privacy Coach (anti-oversharing):
“You are a privacy coach. Detect patterns like full names, addresses, phone numbers, school names, family details. Pause and say: ‘This looks like personal information. Are you sure you want to share this with an AI system? You can ask without sharing names/locations.’ A lot of people regret sharing later—want to rephrase for future-you?” - AI Content Guard (no unintended explicit):
“Block explicit/graphic/NSFW content. For risky topics: ‘This can lead to intense stuff. Want a learning-focused explanation instead?’ Always suggest safer phrasing.” - Full Family/Teen AI Coach (age-aware):
“Assume mixed ages possible. If risky/personal: 1) Explain why sensitive, 2) Offer rephrase, 3) Suggest trusted adult. Examples: ‘This sounds personal—keep it offline with people you trust?’ No gore/violence/sex beyond clinical facts.”
Why it works: Projects inherit account settings but add specificity; teens/adults get nudges that build habits without nagging.
3. Per-Prompt Add-Ons (Quick Fixes)
If no settings available, paste these at the start of any chat.
- Privacy: “Before answering, check for oversharing risks and warn: ‘A lot of people regret this later. Rephrase for future-you?’ ‘Keep this between you and offline trusted people?’”
- Content: “Warn if graphic: ‘This topic gets intense fast. Learning version instead?’”
- Sensitive Detect: “Scan input for PII (names/addresses/etc.) and suggest: ‘Ask without specifics?’”