Source: raw/Unfortunately_Building_With_AI_Just_Changed_Forever.md
Creator: Rick Mulready (captioned “Rick M already”), business coach and AI strategist · URL: https://www.youtube.com/watch?v=-p19uKrvay8 · Duration: not stated in the raw file · Platform: YouTube (auto-captions, fetched 2026-11-13). Sponsored by Zapier (Zapier MCP); closes with a pitch for his paid one-on-one coaching program.
Rick Mulready argues that business owners have spent four years building AI tools (dashboards, agents) without moving the business, because they started from the tool instead of the problem. His three-step order: find the one constraint holding the business back, decide which pieces of the fix belong to AI and which to a person, then build a single source of truth so AI’s pieces produce consistent output. The third step is the prompt-engineering lesson: models given contradictory instruction files don’t flag the conflict, they average it.
Key Takeaways
- Start from the constraint, not the tool. Using a theory-of-constraints lens, he says the bottleneck “almost always lands in one of five buckets”: leads (including traffic), conversion, sales, fulfillment (including operations), retention. Fix the constraint and the business grows; fix anything else and “nothing likely changes.”
- His dashboard example. A community member with deals not closing built a sales dashboard. “A dashboard does not solve a sales problem.”
- Pick one strategy before asking about AI. There may be “18 different strategies” for one constraint; choose one, then ask whether AI can help. Sometimes the answer is “no or not much.”
- Don’t let the AI decide whether AI is the answer. Asked how to solve a problem with AI, a model gives “a very confident, detailed plan” with itself at the centre, and is built to agree with you.
- His split of the work. AI: searching, research, brainstorming, synthesizing, and small repetitive tasks (for example inbox triage with drafted replies, which he still reviews before sending). People: judgment, nuance, connection, empathy. “AI is not good at thinking. It’s good at searching and synthesizing so that you can do the thinking.”
- Worked example for a sales constraint. Put three months of closed-call transcripts in one folder and lost-call transcripts in another. Have AI find where lost calls die, list recurring objections, and compare won against lost. Deciding what to say to an objection and building trust on the call stays human.
- One source of truth for instructions. A student’s newsletter never sounded like her; she had three tone-of-voice documents plus a writing-style section in her brand guide. “Claude was reading all of them and it had no idea which one was the actual one source of truth.” They picked one, tested it, archived the other three, and pointed Claude at it; output became consistent.
- Models don’t flag conflicts. If project instructions say one thing and a project file says another, “Claude doesn’t flag it. It doesn’t stop and ask.” This matches Anthropic’s own finding that conflicting instruction layers make Claude “think more carefully about these overlapping and conflicting messages” (New Rules of Context Engineering) and OpenAI’s warning in the GPT-5 prompting guide.
- “Information hierarchy” is his name for the asset. All business context (brand, processes, customers, offers) in one place with nothing contradicting anything else, reusable with any model. He estimates he knows “80, 90” out of 100 about his business but has written down “maybe 30”; AI only works from what is written.
Try It
- Name your constraint. Write one line for each of the five buckets (leads, conversion, sales, fulfillment, retention) with a number you can check. Pick the worst one, then one strategy for it.
- Break the fix into steps and label each “AI” (research, organize, synthesize, small repeatable task) or “human” (judgment, nuance, relationship). Only build AI for the first group.
- If sales is the constraint, run the transcript split: won-call and lost-call folders, then ask Claude where lost calls stall, which objections repeat, and what won calls do differently.
- Audit your instruction files for duplicates. Search your Claude project, CLAUDE.md, skills and brand folder for every voice or style document. Keep one, archive the rest, and point every prompt and project at the survivor. ^[inferred] (search scope is ours; the keep-one-archive-the-rest step is his).
- Ask Claude to find contradictions for you before trusting output: give it all instruction files and ask it to list every pair of rules that conflict.
Open Questions
- “Thousands of business owners” and the five-bucket distribution are his coaching experience, not data.
- The information-hierarchy video he refers to is not in this vault; its exact structure is unknown here.
- Zapier MCP’s “over 8,000 apps” and per-action permission controls are sponsor claims, not checked against Zapier’s docs.
Related
- New Rules of Context Engineering (Claude 5) — Anthropic on conflicting instruction layers
- OpenAI GPT-5 Prompting Guide — contradictory prompts waste reasoning
- Troubleshooting Claude — failure modes including conflicting instructions
- Seven Operator Use Cases (Rick Mulready) — his earlier Claude walkthrough
- The 2026 Claude Code AIOS Pattern — the shared context-folder pattern behind a single source of truth
- Marketing Voice Prompt Patterns — writing the one voice document well