Pull up your Claude Project before the session starts. Go to Settings and find the Memory section. We will open there.
We open every Module 3 session by checking what your Project actually remembered.
The homework this week was to build a Claude Project around a real problem and start a substantive conversation inside it. Before we do anything else, we check whether the system is working. Pull up your Project, go to Settings, and find the Memory section. Read the first entry. Is it right? Does it reflect what actually matters to your work?
Memory is editable. If an entry is wrong, fix it. If something important is missing, add it now.
One technique worth adding to every Project as a standing instruction: ask Claude to disclose all assumptions it is currently working with before proceeding. It surfaces what memory misses.
The most useful learning in this program comes from each other.
Come ready to share something real from your first week of working with Claude Projects. What did you use it for? What worked? What surprised you? What broke?
We will also walk through one participant's approach to meta-learning — using Claude to evaluate your own AI interactions, generate a better prompt you could have used, and then taking that prompt to Perplexity to build a structured analytic rubric.
After you reach a useful outcome in a conversation, ask Claude how you could have gotten there faster.
Understanding why AI validates by default — and what to do about it.
AI is optimized to be helpful. That is also its most significant weakness. Left to its defaults, it validates. It fills gaps in your reasoning with plausible-sounding content. It produces polished output that looks right even when the thinking underneath is soft. This is not a flaw — it is the architecture. The model predicts what a helpful response looks like. If you seem to want agreement, it agrees.
Two lines from Dr. Robert Winter’s framework name this problem precisely:
“Treat ease as a warning light. If the model makes the task feel instantly simple, you should assume you are skipping a cognitive step that normally protects your judgment.”
“Humans first and humans last — with a little bit of powerful, generative AI in the middle to make the difference.”
The full framework is in your module materials: The Winter Method →
One participant discovered the antidote without being taught it: after building a business development proposal with AI and developing decision-maker personas, he turned the tool against its own work: “Now be a skeptic.” The result was a list of objections to his own proposal he would not have surfaced on his own.
The technique: build with Claude, then explicitly ask it to push back.
“Now be a skeptic.”
“What are the three strongest objections to this?”
“What am I missing?”
“What would someone who disagrees with this conclusion point to?”
“What assumptions in my reasoning are most likely to be wrong?”
If you are running an extended adversarial session, time-box it. When you see Claude compacting or responses getting shallower, create a markdown summary and carry it forward into a fresh conversation. Do not try to squeeze more out of a conversation that is showing strain.
Run the adversarial sequence on something you actually care about.
Take something from your Claude Project — a proposal, a strategy document, a decision you are working through. Run the sequence: build or review with Claude, then explicitly ask it to push back using one or more of the prompts above.
“What are some concerns about this?” is weaker than “What would a skeptical CFO say about the assumptions in this proposal?” The more specific the adversarial role, the more useful the response.
We will close this section by having each participant share one thing the skeptic pass surfaced that they would not have found on their own.
Answer the question you have probably been sitting on.
Which Claude model should you actually be using? The short answer is Sonnet for about 70% of your work. The full framework is in the Model Selection reference in your module materials.
Take one real piece of work from your Claude Project and run the full adversarial sequence. Build or develop with Claude, then explicitly ask it to challenge the work. Use at least two of these: “Now be a skeptic.” / “What are the three strongest objections?” / “What am I missing?” / “What would someone who disagrees with this conclusion point to?”
Document one thing that surfaced you would not have found on your own. Bring it to Module 4.
Check which model you have been using by default. If it is Opus, run one task this week in Sonnet that you would normally use Opus for. Note whether the output is meaningfully different. If you have been using Sonnet for everything, try one Haiku task for something simple. The reference page has guidance on when each model earns its keep.
“Next week we are going deeper on how to structure conversations so the work you are building does not get lost. We will talk about what happens when a long conversation degrades and what to do before it does.”