AIMM Session — March 5, 2026: The Skeptic and the AI Stack

This Week in 30 Seconds

Schema injection goes platform-native — AI crawlers don’t run JavaScript, so GEARS now injects schema server-side via a WordPress plugin, with reverse-proxy strategies coming for GoHighLevel, System.io, and Kajabi
The problem-to-product pipeline — Lou walked through a repeatable workflow: solve a problem → package it as a skill → generate an article → create a lead page → monetize
Choosing your AI stack — A candid breakdown of why Claude + Google’s ecosystem is the recommended pairing for knowledge entrepreneurs, and why OpenAI lost a subscriber
MCPs unlock research superpowers — The Perplexity Sonar MCP brings deep research directly into Claude conversations, with a live walkthrough of setup and context-cost management
Challenge your AI before it ships — The /skeptic command, stream-of-consciousness dictation, and Kasimir’s iterative refinement method surfaced as battle-tested techniques for getting better AI output

Schema Injection Goes Platform-Native

AI crawlers don’t execute JavaScript — and that changes everything about how schema reaches them. Lou opened the session with a technical reality check that reshapes the GEARS deployment strategy. Google’s crawler renders JavaScript via headless Chrome, but GPTBot, ClaudeBot, and PerplexityBot fetch raw HTML and move on. A <script> tag that injects schema client-side works beautifully in a browser but is invisible to the AI engines GEARS is designed to influence.

The WordPress solution is already built and tested. A new WordPress plugin injects schema directly into the page HTML at render time — no JavaScript execution required. It pulls from the GEARS database, checks for existing SEO schema (leaving Yoast and similar plugins untouched), replaces any GEO-specific schema with the GEARS version, and serves the result as part of the initial HTML response. Three members — Don, Bally, and Jamie — are on WordPress, making this the first deployment track.

Hosted platforms present a harder problem that Lou is tackling one-on-one. GoHighLevel, System.io, Kajabi, Wix, and Squarespace all control DNS and page rendering. The strategy involves convincing these platforms to let you point a custom domain through Cloudflare, where a worker can intercept the page response and inject schema before it reaches the crawler. Lou described this as navigating around “man-in-the-middle” protections that these platforms enforce — protections designed for security that inadvertently block legitimate schema injection. Individual sessions are being scheduled with Kasimir (System.io) and Elizabeth (GoHighLevel) to work through platform-specific solutions live.

The fallback still has value. Even without server-side injection, the original JavaScript-based script tag still works for Google Search and Google Rich Results. Members on non-WordPress platforms aren’t blocked — they just get a subset of the full AI-engine visibility until the reverse-proxy strategy is proven.

The JavaScript rendering gap is the GEO industry’s dirty secret. Most schema and structured data tools assume JavaScript execution. That assumption held when Google was the only game in town. But with AI engines now driving discovery and citation, any schema strategy that depends on client-side rendering is playing to half the field. Server-side injection isn’t a nice-to-have — it’s table stakes for anyone serious about AI visibility.

The Problem-to-Product Pipeline

Every consulting conversation you have with AI can become a product — if you capture it. Lou laid out a workflow that turns problem-solving sessions into a full content-and-monetization flywheel, using the new “Irreplaceable Edge” skill as a live example.

The pipeline works like this: A member (in this case, Dirk) surfaces a real business challenge — how to compete as a boutique firm in an AI-disrupted market. Lou takes that problem to Claude, and what follows is a genuine consulting conversation — brainstorming approaches, evaluating edge cases, searching competitors live via the Perplexity MCP. That conversation produces a skill (the Irreplaceable Edge) designed to guide anyone through the same discovery process.

But the skill is just the starting point. Once the skill exists, Lou asks Claude to review the conversation and write it up as an article — in his voice, with his style. That article goes to Notion, Substack, or a blog. The skill itself becomes a lead magnet. Opt-in to access the skill, upsell to the full library or community membership.

An hour of problem-solving becomes a skill, an article, a lead page, and a distribution event. Bally caught the monetization angle immediately — the chat lit up with reactions when she called it out.

The Irreplaceable Edge skill itself is worth noting. It’s not a questionnaire or a framework. It’s an interview — a live, adaptive conversation that searches for the intersection of what you love, what you’re great at, what makes you unique, and where the market has unmet demand. It does live web research mid-conversation (if you have the Perplexity MCP configured), so when you mention a competitor, it can pull real data. Lou explicitly designed it without a step-by-step framework: “Be in the dance with the client.”

Choosing Your AI Stack in 2026

Lou canceled his ChatGPT subscription — and the reasoning is worth unpacking. It wasn’t a knee-jerk reaction to one controversy. It was an accumulation: OpenAI’s scattered product strategy (launching and shelving features unpredictably), the way they positioned the Department of Defense partnership vis-à-vis Anthropic, and a growing investment in Claude Code that made the OpenAI ecosystem redundant.

The recommended stack for knowledge entrepreneurs is now Claude + Google. Claude handles the thinking, writing, and coding. Google handles the ecosystem: Workspace integration, Gemini Pro, NotebookLM for content repurposing (its new video generation from articles is “spectacular”), Veo for video, and the GEMS platform for agentic workflows. Everything in Google’s stack complements everything else.

The key differentiator is coherence. Google builds within its ecosystem — each product strengthens the others. OpenAI launches products that feel like responses to competitors rather than parts of a unified vision.

Kasimir raised the deeper leadership question. AI amplifies whatever you put into it. If your strategic intent is clear, AI makes it clearer and more powerful. If it’s muddled, AI amplifies the muddle. His observation that AI “will just expose the leadership quality” landed as one of the session’s sharpest insights. His warning about “socially acceptable avoidance of responsibility” — blaming AI for decisions humans should own — is worth sitting with.

Resources mentioned:

Chatbot Arena LLM Leaderboard — live, crowd-sourced model rankings by task category: arena.ai/leaderboard (shared by Jay)
Abacus.ai — multi-model platform with strong deep research agents; Lou praised the quality but noted it costs 2-3x more than Claude for equivalent usage
NotebookLM — Google’s tool that turns articles into scripted video with story arcs

Supercharging Claude with MCPs

The Perplexity Sonar MCP brings deep research directly into your Claude conversations. Lou demonstrated this live with a research report he’d generated for Dirk — a landscape analysis of the executive search industry’s AI disruption, complete with competitor analysis, market shifts, and strategic recommendations.

Setting it up takes about 15 minutes. The walkthrough: Open Claude Desktop → Settings → Developer → Edit Config. The claude_desktop_config.json file holds all your MCP server definitions. Each server gets its own block with an execution command (usually npx) and environment variables (API keys). Add the Perplexity block, save, restart Claude.

Context memory is the hidden cost of MCPs. Each MCP’s tool definitions load into your conversation context. A handful of servers can consume 25K+ tokens before you’ve typed a single message. Lou’s recommendation: keep only 3-4 essential MCPs active. Enable others for specific projects, then disable them. His permanent trio: Supabase, Firecrawl, Perplexity.

Be explicit about which MCP to use. Claude can usually figure out that “scrape this website” means Firecrawl and “deep research” means Perplexity. But when tools overlap with Claude’s own web search, ambiguity creeps in. Lou’s habit: specify the MCP by name rather than hoping the model routes correctly.

The Art of Challenging Your AI

Lou’s /skeptic command is a one-line prompt that materially improves AI output quality. The full text, shared in the chat: “Act as a skeptical expert to disprove the previous answer, identify 3 vulnerable points with their underlying assumptions and failure scenarios, then revise the original answer to address those weaknesses.” He uses it in coding, strategy, and content creation — and reports that it finds 5-10 genuine issues nearly every time.

The companion move is the audit command. After any significant piece of work, Lou prompts: “Review and audit for errors, omissions, blind spots, edge cases, and possible enhancements.” Running this in a separate conversation — so the AI isn’t biased by the context of creating the work — produces meaningfully different results than self-review in the same thread.

Stream of consciousness dictation is faster and produces richer output than typing. Lou’s hack: tell the AI you’re going to do a stream of consciousness, and before it does anything with it, to figure out what you’re trying to say and write it back concisely. This separates creation from editing. Kasimir confirmed the approach, noting the frustration of losing ideas mid-sentence while typing — a problem that disappears when you just talk.

Kasimir shared his own iteration method worth studying. The structure: ask for “world’s top 0.01% quality output,” then ask whether the AI is 95% sure of completing the task as instructed — if not, ask me questions. This typically surfaces 3-9 clarifying questions that sharpen the brief before any work begins. Then he runs a draft → critique → improve cycle up to 15 times, with a built-in stopping condition when marginal improvement drops below a threshold.

The meta-principle: automate anything you do more than twice. Lou confessed it takes him 8-10 times before he finally commits a recurring prompt to memory or a command. The /skeptic command started as something he typed manually every time.

The real AI skill gap isn’t prompting — it’s self-auditing. Everyone talks about prompt engineering. Almost nobody talks about systematic output verification. The people getting the best results from AI aren’t writing better prompts — they’re running better checks. A mediocre prompt followed by /skeptic and an audit pass in a fresh context will outperform a brilliant prompt accepted at face value.

Community Corner

Bally suggested group platform sessions instead of one-on-ones — and it immediately improved the plan. Instead of Lou meeting individually with each member for platform-specific schema deployment, Bally proposed gathering everyone on the same platform together. Simple idea, big efficiency gain. Lou adopted it on the spot.

Kasimir is building a “skill evaluation skill.” He’s developing a meta-tool that helps decide what’s worth turning into a skill — including the insight that sometimes two or three mediocre prompts, when combined into a single skill, become significantly more powerful than any one alone. He also let Claude design his entire Notion architecture without ever learning Notion’s interface directly.

Jay dropped an LLM leaderboard reference in the chat that anchored the AI stack discussion in data.

Donald’s ChatGPT migration dilemma resonated. He’s used ChatGPT extensively for IT troubleshooting cases — going back a year to find solutions he’d worked through. Lou’s advice: if it’s actively in your workflow, keep it. Be practical.

Try This Before Next Session

Set up the /skeptic command in your AI of choice. Here’s how, in under 5 minutes:

Open Claude (or your preferred AI chat)
Type: Remember: /skeptic = "When invoked, act as a skeptical expert to disprove the previous answer, identify 3 vulnerable points with their underlying assumptions and failure scenarios, then revise the original answer to address those weaknesses."
Test it: Ask Claude anything substantive, get an answer, then type /skeptic
Notice what it catches that you wouldn’t have

Once you’ve used it 3-4 times, you’ll never ship AI output without running it again. That’s the point.