A demo that works one time in ten is a liability, not a feature. Here is why LLMs hallucinate, why their confidence scores lie, and how I validate AI output in WordPress before it ever reaches a user.
I stopped hand-reviewing every line my coding agents write. What I do instead: fine-grained permissions in settings.json, a strict deny list for destructive commands, and a second agent that reviews the first one’s work.
I spent months trying to make Claude Code my main tool, but since GPT-5.5 I’ve moved complex WordPress refactoring back to OpenAI Codex. Here’s my setup: reasoning levels, Playwright MCP, and a worktree alias.
WordPress 7.0 brings a central Connectors hub and a Connector Approval flow for AI credentials. Here’s what the May 2026 plan means in practice: custom tables for request logging, the AI Plugin 1.0.0 release, and what the approval flow does and doesn’t protect.
WordPress 7.0 “Armstrong” puts an AI Client and a provider-agnostic Abilities API into core, registers blocks from PHP alone, and raises the minimum PHP version to 7.4. Here is what changes if you maintain client sites.
Chunk-and-embed RAG loses the hierarchy that gives a 200-page contract its meaning. How the Proxy-Pointer framework indexes structural breadcrumbs instead of raw text, and how the two-stage retrieval works in Python.
Notes from the core performance chat the day before WordPress 7.0 ships: on-demand block style loading for classic themes, Performance Lab moving to PHP 7.4, npm ignore-scripts, and an FTP filesystem install fix.
MCP servers are becoming the plugin bloat of the AI era. Why a sandboxed CLI gives an agent more capability per token than a menu of narrow tools, and how to keep shell access from being a liability.
I stopped hand-holding Claude Code and gave it a nightly review instead: a self-reflection skill, a 2 AM cron job, and a rules file it edits itself. This is the setup I run on real WordPress work.
Every AI harness keeps its own memory, so moving from Cursor to Claude Code means starting from zero context. Here is the shared layer I build instead: lifecycle hooks writing into Neo4j, plus the batch job that turns raw events into notes.