How to prompt agents that can write to your data: split, list, confirm
SendPulse · reddit · 2026-10-23
SendPulse shares prompt patterns for agents connected via MCP (Claude, ChatGPT, Cursor) that can create records and update statuses—not just answer questions. Failures are rarely wrong answers; they're sensible actions applied to the wrong records (e.g. the agent's definition of "this week"). Key rules: 1) split instructions into find-then-write with an explicit confirmation step and full row list before anything is permanent; 2) be explicit about what qualifies instead of vague phrases like "this week's orders"; 3) ask for the actual rows, not "I found 23 matches"—rows reveal whether the agent understood; 4) enable fewer tools than you think you need, since every enabled tool is an action that can happen. Author admits the confirm step means reading lists all day and questions whether it scales.
More from coding & agent
- Zi Wei Dou Shu open-source astrology engine trends on GitHub with 6,565 stars — Renhuai123 · 2026-10-23
- One ComfyUI node replaces 42-step MiniMax H3 ref-to-video workflow — Dudelydad78 · 2026-10-23
- opencode model catches and rejects a prompt injection mid-reasoning — KomithErrant · 2026-10-23
- Voice AI that sounds human but can't handle interruptions isn't production-ready — djgoel123 · 2026-10-23
- Claude checks the Mac's battery before saying goodnight — "plug in the charger first" — kingslyj · 2026-10-23
- Give Claude 17M+ films and 14M audio recordings for $0 via Internet Archive API — _AustinCalvert_ · 2026-10-23