Two assignments. Real work from your Project. Bring results to Module 4.
Take one real piece of work from your Claude Project — a proposal, a strategy document, a decision you are working through — and run the full adversarial sequence.
Get to a result you are reasonably satisfied with.
Use at least two of the following prompts:
Something you would not have found on your own. Bring that finding to Module 4. One sentence is enough: what did the adversarial pass catch that you missed?
Specificity matters more than the prompt itself. “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.
Check which Claude model you have been using by default.
Run one task in Sonnet that you would normally use Opus for. Note whether the output is meaningfully different for that type of work.
Try one Haiku task for something simple — an email draft, a quick summary, a classification task. Note whether Haiku handled it well.
Check the model selector in the top left of any Claude conversation.
The reference page has the full framework for when each model earns its keep.
Model Selection Reference →If your adversarial session runs long and you see Claude compacting the conversation, that is your signal to stop and capture. Create a markdown summary of what you have built together and carry it forward into a fresh conversation or drop it into your Project files. The work is not lost as long as you capture it before the conversation degrades.
Module 4 opens with a question: “Who has had a conversation go sideways in the last few weeks?” This case study is the setup. Read the “What Broke” section before the session.
→ When Conversations Break Down The failure section of a real multi-model workflow. What went wrong, why it went wrong, and what the recovery looked like.