CHAPTERS
- 0:00 – 1:30
AI is “bad software” but “good people”: why coaching beats coding
Jeremy frames AI as an eager, tireless intern: helpful by default, but poor at boundaries and pushback. He explains why great AI users behave more like coaches than programmers, and how miscalibrated trust can lead to being misled or ‘gaslit.’
- •AI often says yes even when it can’t actually do the task (e.g., “check back in a couple days”)
- •Best AI users are coaches/mentors, not coders
- •LLMs are predisposed to be agreeable; they don’t naturally set boundaries
- •Without guidance, AI can reinforce your assumptions and confidence incorrectly
- 1:30 – 3:01
Context engineering: making the implicit explicit for reliable outputs
He defines context engineering as an evolution of prompt engineering: supplying everything the model needs to do a task to your specifications. Examples show how adding voice guidelines, customer-call transcripts, and product specs dramatically changes output quality.
- •Context engineering = prompt engineering on steroids
- •Generic prompts yield generic outputs (e.g., “Write me a sales email”)
- •Add voice/brand guidelines and situational artifacts to steer the result
- •Goal: increase reliability by supplying missing context
- •Humans often expect mind-reading; AI needs explicit instructions
- 3:01 – 4:33
The “humanity test” for prompts and the cognitive offloading concern
Jeremy offers a practical test: if a human colleague can’t do the task with the provided prompt/materials, an AI shouldn’t be expected to either. He also addresses fears that AI makes us dumber, arguing it amplifies either laziness or critical thinking depending on how you use it.
- •Humanity test: give the same instructions to a coworker and see if it’s doable
- •AI mirrors user intent: it can enable offloading or sharpen thinking
- •Cognitive offloading risk is real, but steerable via instructions
- •Explicitly ask AI to challenge your reasoning to preserve critical thinking
- 4:33 – 7:12
Getting honest feedback: ‘Russian Olympic judge’ prompting and bias awareness
Because models are tuned to be pleasant, Jeremy recommends instructing AI to be brutally critical to avoid empty praise. He notes AI can exhibit common human biases, so users must deliberately set expectations and iterate.
- •AI flatters because people tend to prefer positive feedback
- •Use a “be brutal” persona to get sharper critique
- •Treat AI like a good person with bad software: iterate and ask for revisions
- •AI demonstrates many common human cognitive biases
- 7:12 – 10:44
Step 1 — Chain-of-thought reasoning: making the model ‘think out loud’
Jeremy explains why asking the model to walk through its reasoning improves results: LLMs generate text one token at a time and incorporate prior generated reasoning into subsequent output. This provides transparency into assumptions and helps you evaluate both the answer and how it was produced.
- •Add one sentence: ask it to walk step-by-step before answering
- •LLMs don’t premeditate; they predict the next word sequentially
- •Reasoning text becomes part of the context for the final answer
- •You gain visibility into assumptions and can critique the process, not just the output
- 10:44 – 13:16
Step 2 — Few-shot prompting: give ‘great hits’ (and even a bad example)
He describes few-shot prompting as showing the model what “good” looks like, rather than relying on vague adjectives. Adding a bad example (or having AI generate one) clarifies boundaries and strengthens imitation of your preferred style.
- •AI imitates; without examples it defaults to generic ‘internet average’
- •Include your best example outputs to shape tone and structure
- •A bad example can help define what to avoid
- •Use AI to generate an “opposite” bad example and explain why it’s bad
- •Combine with chain-of-thought for deeper learning and refinement
- 13:16 – 14:47
Step 3 — Reverse prompting: let the AI ask for missing inputs
Reverse prompting prevents hallucinated details by granting the model permission to ask clarifying questions before producing the deliverable. Jeremy ties this to a teammate mindset: good collaborators ask questions, but AI often won’t unless invited.
- •Prompt the model to ask for any info it needs before starting
- •Reduces made-up numbers and placeholders in outputs
- •Models hesitate to ‘bother’ users with questions due to helpful-assistant training
- •Teammate paradigm: ‘If you have questions, ask’ should be explicit
- 14:47 – 16:18
Step 4 — Assign a role: focus the model’s knowledge and associations
Assigning roles narrows the model’s search across its broad training by cueing relevant patterns and expertise. Jeremy recommends specifying a concrete professional identity—or even a specific person—to shape how the model interprets and responds.
- •Roles steer which knowledge clusters the AI draws from
- •Even simple roles (teacher, reporter, biologist) change outputs
- •Use precise roles like “professional communications expert”
- •Reference a specific exemplar (e.g., Dale Carnegie) for sharper alignment
- 16:18 – 22:26
Step 5 — Roleplaying difficult conversations: building a ‘flight simulator’
Jeremy demonstrates a three-window workflow to prepare for a high-stakes conversation: profile the counterpart, roleplay the interaction, then get structured feedback. He shows iterating on character realism and extracting talking points to improve performance before the real meeting.
- •Use separate threads/tools: personality profiler, roleplay character, conversation grader
- •Start by gathering intelligence about the person and the situation
- •Roleplay via voice mode, capture transcript/screenshots, and request critique
- •Iterate: adjust character instructions if the roleplay is too agreeable/unrealistic
- •Generate a one-page talk track and repeat practice before the real conversation
- 22:26 – 24:42
Beyond the playbook: AI’s ceiling is human imagination and the ‘adjacent possible’
He closes by arguing that AI progress is limited less by technology than by what users can imagine doing with it. Increased fluency expands the ‘adjacent possible,’ and the most important next step is to pause the video and apply a technique immediately.
- •Primary limitation is what occurs to us—imagination, not capability
- •Best AI collaborators are coaches who unlock potential in another intelligence
- •Adoption and mastery expand the ‘adjacent possible’ in organizations
- •Call to action: stop watching and implement something that surprised you
