Sorry mate, but I’m surprise that you didn’t catch the fake engagement within that community. 😞 It saddens me that you spent so much time and effort replying to the few actual/real members. I’m confused, however, as to why the Skool platform removed you if you were working with @Jay Feldman? 🤔 I’m also confused as to what your copy has to do with a suspended FB account? 🤔 … but, I digress. I appreciate your story for sure. The TL;DR is that Metrics can be gamed. Relationships can’t. Fake engagement might win the leaderboard for a minute, but real trust is what survives when accounts vanish, algorithms change, or the manufactured activity dries up. Build the community people still value after the platform stops rewarding the shortcut.

@Ian Kirk I find that I’m helping (or attempting to at least 🤣) more than I’m learning 😊

Deets? 🤔😁

Controlling CC is as simple as simplifying the chunks that it is processing at any one time. You define each activity/chunk like you would a task on your project dev timeline. You also ask it to write outputs for each phase to “outputs-phase0.md” as to what it did, and how the test results fared against the criteria that you set. Then, esit clause.md to ensure that it calls each of the phased outputs each time it works on the next phase (linear consistency).

I run my own business on this very thing, and I’m currently building Australia’s first sovereign AI brain for mid-tier legal firms, so I absolutely know how to build a “self-contained” AI brain. The simplest framing is that you need to train the corpus (knowledge base) on the customer’s journey, retrieve answers from high-quality source material, and evaluate it against real electrician customer questions. That gives you an AI that can answer, educate, qualify, and convert, without sounding like generic chatbot sludge. You’ll need to use RAG, retrieval augmented generation, versus “training a model”, to build a structured knowledge base, retrieve the right source material at answer time, then make the AI answer *only* from that source material. For electricians, I’d structure the knowledge base around customer intent and buying stage, rather than just dumping documents into the model’s project folder. You’ll need to develop some basic layers, such as audience, problem, journey, conversion, and proof. Here is the basic practical build process: 1. Collect source material, website pages, service pages, call scripts, sales emails, quote templates, FAQs, reviews, job notes, case studies, local SEO pages, warranty language, and compliance explanations. 2. Clean it, remove duplicates, outdated prices, weak AI content, and anything that sounds generic. 3. Tag every item with metadata, audience, service type, journey stage, urgency, location, objection, offer, proof type, and preferred CTA. 4. Chunk by meaning, not arbitrary length. Keep chunks focused on one question, one service, one objection, or one buying-stage problem. 5. Write answer templates for each use case, for example emergency call response, quote follow-up, website FAQ, Google Business Profile response, email nurture, ad landing page, and chatbot answer. 6. Build an eval set, 50 to 100 real customer questions, then score each answer for accuracy, helpfulness, citation/source use, safety, brand tone, and conversion strength. You need custom criteria for question-answering. 7. Improve retrieval before blaming the model. Poor answers usually come from weak source content, bad chunking, missing metadata, poor retrieval, or no eval loop.

Depends on what you are doing with CC and how complex. Dropping the model to an earlier version will get you further, especially if you don’t need the complex thinking of the latest models. Im also curious as to your workflow, and how much you are asking CC to orchestrate within the workflow, because that could be the issue for your excessive token burn.