> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.meetgail.com/platform/communications/actions-creation-guide/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.meetgail.com/_mcp/server. # Actions Creation guide > A best practices guide for you to create the best Actions possible # How to Build Bulletproof Actions ## 1. Core Mindset * Treat the computer like a **super-intelligent baby** * Never assume context, intent, or understanding * If something feels “obvious” to a human, it is **not obvious** to the system * Prompting and scripting = **coding in plain English (pseudocode)** --- ## 2. Think in Extreme Granularity * Every step must be broken into **explicit, literal steps** * Never rely on implied behavior Ask yourself: * *How does it know where to get this?* * *How does it know what to say vs what to do?* * *What happens if the caller says something slightly different?* --- ## 3. Actions Structure (Always) Every use action clearly contain: ### Title * Naming only * Has zero impact on performance ### Objective * Defines *when* the action should run * Must be a **conditional statement** * Never list steps here * Answers: *“Under what condition does Gail enter this flow?”* * Think: *“If the caller wants (x, y, z), then follow these steps:”* ### Steps * Step-by-step instructions * Linear * Explicit * Granular * No ambiguity --- ## 4. Conditionals Are Gates (Not Suggestions) Use conditionals to control flow precisely. * **If / Then** * One condition must be true * **Or** * Any one condition can be true * **And** * All conditions must be true * **Nor** * None of the conditions can be true Conditionals decide **which path the flow follows**. --- ## 5. Directives vs Spoken Steps (Critical) ### Directives * Tell Gail what to do internally * Gail should **never say these out loud** Examples: * Show empathy * Do not disclose timing * Do not mention internal processes ### Spoken Steps Must explicitly start with: * “Ask the caller…” * “Tell the caller…” Never mix directives with spoken text — this breaks realism. --- ## 6. What Bad Scripts Have in Common * Objectives that describe steps instead of conditions * Vague steps like: * “Guide the caller…” * “Provide the link…” * Missing answers to: * Where does the info come from? * How is it delivered? * What exact questions are asked? * Transfers without determining **who** or **why** --- ## 7. What Good Scripts Do Well * Objectives clearly define **when the flow starts** * Steps are: * Linear * Explicit * Unambiguous * Conditional logic is used **inside steps** to handle variations * One use case can handle multiple scenarios if: * The end outcome is the same * The logic paths are clearly gated --- ## 8. Always Think Linearly * Actions are **flowcharts written in words** * Gail checks steps **top to bottom** * If a condition isn’t met → move to the next step * Once a condition is met → execute → exit flow --- ## 9. Start With the Outcome, Then Work Backwards Ask: * What is the final outcome? (transfer, quote, payment, info, etc.) * What conditions must be satisfied to reach that outcome? * What info must be gathered **before** taking steps? * What decisions change the path? --- ## 10. Design Like You’re Teaching a Child Ask yourself: * Would a 5-year-old know what to ask next? * Would they know the order? * Would they know what *not* to say? * If not → add steps --- ## 11. Key Rules to Remember * Be specific * Be literal * Be granular * Avoid assumptions * One step = one clear instruction * Linear > clever * Clarity beats brevity > Documentation for Gail, the AI platform for financial services. Learn how to set up GailGPT and Gail Agent to automate customer communications for insurance, banking, and finance.