Skip to main content
ZeroWidth

Office hours

Pete builds with ZeroWidth live on Twitch. Every stream is here, edited down, with chapters to jump straight to the part you want.

Subscribe on YouTubeWatch live on Twitch

Oct 7, 2026 · 46 min · Prism · Watch on YouTube

Exploring a game where every NPC is an LLM

Pete uses Prism to explore a game where every character is a language model, checks whether the idea already exists, then writes and takes a survey with AI follow-up questions.

  1. 0:00Yesterday's recap
  2. 0:44A game where every NPC is an LLM
  3. 2:27What Prism is for
  4. 3:27Pushing an idea along an axis
  5. 7:19Generated images, and the flows behind Prism
  6. 8:25When your product outruns your testing
  7. 10:06Exploring without a plan
  8. 11:17What only works if the NPC is a model
  9. 15:48The ideas between two ideas
  10. 18:30Making it a web game
  11. 20:54When the model loses the plot
  12. 23:29Each guest is its own model
  13. 26:03Desk research: does it already exist?
  14. 28:49Not a new idea
  15. 29:26Drafting a survey with zv1
  16. 31:24AI follow-up questions
  17. 33:59Taking the survey
  18. 38:04Results and the printable report
  19. 39:25Research that calls you out
  20. 40:26Clustering the field into themes
  21. 42:00Sharing the field
  22. 43:17Quotas and automatic re-analysis
  23. 44:21Research on a schedule
  24. 45:07What's next

Oct 6, 2026 · 57 min · Watch on YouTube

Does your AI get better the more people use it?

Pete builds a texting-style agent in Workbench, measures it in Caliper, improves it with zv1 and adds memory, to show which kinds of better happen on their own and which ones you have to design.

  1. 0:00Does it get better the more you use it?
  2. 0:42Brand kits and on-brand decks in Napkin
  3. 3:39Three things people mean by better
  4. 4:29Building a texting-style agent in Workbench
  5. 10:27From anecdotes to evals
  6. 11:53Datasets and rubrics in Caliper
  7. 15:27Generating test cases with zv1
  8. 18:33Using the MCP server from Claude Code or Codex
  9. 19:43Why multi-turn evals matter
  10. 23:23When the judge misses tone
  11. 25:05A tone rubric: 88% to 74%
  12. 28:46Why the number matters
  13. 32:23Improving the agent with zv1
  14. 34:44Keep test cases out of your prompt
  15. 35:57Up 21 points
  16. 37:50Using an AI doesn't train it
  17. 38:58Shared memory or memory per person
  18. 40:39Profile data in the system prompt
  19. 42:20How the agent writes to memory
  20. 44:32Memory is retrieval
  21. 45:08Debugging a memory that won't stick
  22. 48:31Deciding what gets remembered
  23. 49:32So, did it get better?
  24. 50:07Evals for memory and personal preferences
  25. 53:15Prompt injection through a profile name
  26. 54:40None of this is machine learning

Try ZeroWidth on your own work. It's free to start.

Create a free accountJoin office hours live on Twitch