AI generated dialog research

Project documents start here · methodology · system prompts

Start here

The three schemes how a dialogue can be generated: two brains and a robot, one actor in many disciplines, or one writer scripting both sides
Cost analysis — study 001 what each scheme actually cost; normalised tables (Opus, Sonnet); architecture vs model as cost levers
Project conventions (CLAUDE.md) what the project is, file structure, working norms
README public-facing overview

Methodology

How to write character profiles profile schema, the single-source rule, the improvisation policy, productive-mode for interviewers
How to write scenes per-actor briefs, off-limits vs. don't-lead-with, what each actor knows vs. doesn't
How to orchestrate multi-agent dialogue the three schemes in spec form, dumb-pipe rules, close detection, event log format
How to audit the seven checks, the audit procedure, known failure shapes from prior experiments

System prompts

Actor the performer's stable posture; thinking-block rules; closing the scene
Coordinator the LLM coordinator's dumb-pipe discipline (scheme 2)
Monolithic the single-LLM dialogue generator (scheme 3 — one writer scripting both sides)
Auditor the auditor's posture and what to refuse to do

Experiment cards require JavaScript to load (they're rendered from experiments.json). The folder experiments/ contains every run if you want to browse directly.