What it is
synthkit is a command-line kit for trying out words before showing them to people. It builds a cast of simulated personas from a brief of facts you supply, then shows them taglines, headlines or offers and records how each one reacts. It is careful about what it claims: results are good for ranking options against each other and surfacing objections, not for predicting real numbers. Each report ends with a list of things to check with actual people.
How it works
- Start a project and write a grounding brief of facts about the people you want to reach.
- synthkit proposes audience segments; each must point to a fact in the brief.
- It generates personas for each segment, each with an attitude dial toward AI from hostile (-2) to enthusiastic (+2).
- Personas are quizzed to check they stay consistent.
- They react to your messages alone, in a focus group, or in an interview. A deliberately weak message is included as a control.
- Reactions are extracted into a fixed format and summarized by segment and attitude, with a reality-check list.
What exists today
- Five commands covering the full path: new project, audiences, personas, test, report
- Prompt templates for each step
- A cache of every AI call, so reruns are cheap, and separate cheap draft and better final model settings
- Reports as Markdown, CSV and JSON, plus a transcript for every persona
- Ten completed runs on one internal project
What isn't there yet
- Only used on one project so far
- No web interface; command line only
- The JSON mode with OpenRouter is still marked experimental
- No public sample run on a neutral topic to show people
- Run folders get large because of the AI-call cache
Under the hood
- Python, built on Microsoft's open-source TinyTroupe persona simulator, pinned to a fixed version
- Talks to AI models through OpenRouter or any OpenAI-style service
- Projects are folders of YAML and Markdown files
- Every reaction is pulled out with a fixed schema so results can be compared across runs