Dev builds an AI webmaster on Grok Bot that non-technical clients email directly
beetz12 · reddit · 2026-09-26
A developer maintaining a small nonprofit's site realized he was just routing his non-technical client's requests into a coding agent, so he built "Ivan": an AI webmaster on xAI's Grok Bot (each bot gets a persistent cloud computer with terminal, browser and files) that the client contacts via plain email.
Core design:
- Docs are the webmaster: guardrails (never edit production, never touch payments without him), a services inventory, a decision log, a persona, and a rebuild guide all live in the repo, keeping the bot itself replaceable.
- One pipeline: acknowledge with ETA → branch → private preview link → logged approval → merge and verify live.
- Three lanes: content changes Ivan handles alone; small code changes need regression tests plus sign-off from a second bot that reads the diff cold; money or personal data always escalates to a human ticket.
- Safety nets: scheduled health checks, a dry-run rollback script, and auto-restore after three failed fix attempts.
Results: an intentional fire drill surfaced two real issues before launch; day one delivered 7 flyer revisions in under 4 hours. But Ivan once misread a blurry photo as the client's, and reported new text "fit cleanly" when it was visibly smaller with old-line remnants—prompting a standing rule: when the agent says it's done, open the file.
Lessons: write instructions into a versioned prompt file in the repo, not chat; templatize the setup (he stood up a second webmaster for another site); and give every agent on a shared machine its own SSH key after a new agent silently overwrote the old one's key.
More from coding & agent
- 4D docs MCP server brings cached command documentation to AI coding assistants — modelcontextprotocol · 2026-09-26
- AI agents ranked by build difficulty: computer-use hardest, email agents easiest — sahilypatel · 2026-09-26
- System 1 models return typed answers in one encoder pass: 45ms per query on a laptop CPU, but often confidently wrong — ghumare64 · 2026-09-26
- BrainAPI Open-Sources a Knowledge Graph–Powered Memory Layer for AI Agents — adnan_hashmi · 2026-09-26
- Your agent didn't break — your delegation model did: the agent-permission problem — Fantastic-Sleep-3352 · 2026-09-26
- 128GB local LLM server forges 80M tokens a month on just $6 of electricity — No-Fuel-9202 · 2026-09-26