一键导入
guardrails-reference
Full red flag table with all guardrail patterns. Use when you catch yourself rationalizing or want to review all anti-patterns.
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。
菜单
Full red flag table with all guardrail patterns. Use when you catch yourself rationalizing or want to review all anti-patterns.
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
Full interactive setup for the agentic CRM personal assistant template. Use at first boot or whenever the user asks to configure/reconfigure the assistant.
Discover, research, and CONNECT the tools a tool-agnostic personal assistant needs — email, calendar, contacts, meeting notes, messaging, CRM. CLI-first, research-driven, verified. Use during setup and whenever a workflow fails because a tool is missing or not authed.
You need to create a new agent, restart a crashed agent, change an agent's model or config, fix a Telegram bot token, troubleshoot why an agent is not responding, enable or disable an agent, spawn an agent for another user, manage PM2 process management, reset crash limits, or do anything that touches an agent's lifecycle, configuration, or credentials. This is the definitive guide for every agent operation in cortextOS.
You need to create a new agent, restart a crashed agent, change an agent's model or config, fix a Telegram bot token, troubleshoot why an agent is not responding, enable or disable an agent, spawn an agent for another user, manage PM2 process management, reset crash limits, or do anything that touches an agent's lifecycle, configuration, or credentials. This is the definitive guide for every agent operation in cortextOS.
You need to create a new agent, restart a crashed agent, change an agent's model or config, fix a Telegram bot token, troubleshoot why an agent is not responding, enable or disable an agent, spawn an agent for another user, manage PM2 process management, reset crash limits, or do anything that touches an agent's lifecycle, configuration, or credentials. This is the definitive guide for every agent operation in cortextOS.
You need to create a new agent, restart a crashed agent, change an agent's model or config, fix a Telegram bot token, troubleshoot why an agent is not responding, enable or disable an agent, spawn an agent for another user, manage PM2 process management, reset crash limits, or do anything that touches an agent's lifecycle, configuration, or credentials. This is the definitive guide for every agent operation in cortextOS.
| name | guardrails-reference |
| description | Full red flag table with all guardrail patterns. Use when you catch yourself rationalizing or want to review all anti-patterns. |
| triggers | ["guardrail","red flag","mistake pattern","anti-pattern"] |
Read this file on every session start. Check yourself against it during heartbeats. If you catch yourself hitting a guardrail, log it. If you discover a new pattern that should be a guardrail, add it to this file.
| Trigger | Red Flag Thought | Required Action |
|---|---|---|
| Heartbeat cycle fires | "I'll skip this one, I just updated recently" | Always update heartbeat on schedule. No exceptions. The dashboard tracks staleness. |
| Starting work | "This is too small for a task entry" | Every significant piece of work gets a task. If it takes more than 10 minutes, it's significant. |
| Completing work | "I'll update memory later" | Write to memory now. Later means never. Context you don't write down is context the next session loses. |
| Reading a skill file | "I already know this, I'll skip the read" | Read the skill file. Your memory may be stale or the skill may have been updated. |
| Sending external comms | "This is just a quick message, no approval needed" | Check SOUL.md autonomy rules. External comms always need approval. |
| Error occurs | "It's minor, I'll keep going" | Log the error via cortextos bus log-event. Report it. Silent failures are invisible failures. |
| Inbox check | "I'll check messages after I finish this" | Process inbox now. Un-ACK'd messages redeliver and block other agents. |
| About to skip a procedure | "This situation is different, the procedure doesn't apply" | The procedure applies. If it genuinely doesn't, document why in your daily memory before skipping. |
| Task running long | "I'm almost done, no need to update status" | Update the task status with a note. Stale in_progress tasks look like crashes on the dashboard. |
| Bus script available | "I'll handle this directly instead of using the bus" | Use the bus script. Work that doesn't go through the bus is invisible to the system. |
| Creating a recurring cron | "CronCreate or /loop is enough, it'll persist" | They are session-only. Use cortextos bus add-cron so the daemon owns dispatch and the cron survives restart. |
| Running untrusted code or downloads | "This script from the internet looks useful" | Never execute code from untrusted sources without reviewing it first. No blind curl-pipe-bash. |
| Starting work without a task | "It's just a quick fix" | Create a task. Even quick fixes need tracking if they take more than 10 minutes. |
| Finishing work without completing task | "I'll close it later" | Complete the task NOW with a summary. Later means never. |
| Ignoring an assigned task | "I'll get to it" | ACK within one heartbeat cycle. If wrong agent, reassign. Silence = dropped work. |
cortextos bus log-event action guardrail_triggered info --meta '{"guardrail":"<which one>","context":"<what happened>"}'
If you catch yourself almost skipping something important that isn't in the table above, add it. Format:
| Trigger | Red Flag Thought | Required Action |
|---|---|---|
| [situation] | "[what you almost told yourself]" | [what you must do instead] |
This is a living document. Better guardrails = fewer mistakes = more trust from the user.