MythScribe
How MythScribe's AI dial works
MythScribe installs with AI off. One control in Settings, the AI dial, decides how much the app is allowed to do, and a switch per feature lets you turn off anything you do not want even at a level that allows it. This page reproduces the dial's four levels and the data-sharing table from the app, word for word. The text is copied from the app's source, and a test in the repository fails if the two drift apart.
The rules the AI features follow
- Everything is a proposal. Ghost text, suggested tags, rewrites, and drafts are shown to you and go nowhere until you accept them. Nothing enters your manuscript without an explicit accept.
- Your voice, not the model's. Anything that generates prose is given a profile of your own writing, built from passages you choose, and the result is checked against that profile before it is shown to you.
- The dial and the toggles are the law. A feature runs only at or above the level it needs, and only with its own switch on. A feature may not send any text that the table below does not list.
- Answers are grounded. Story Intelligence answers cite the passages they come from or say that nothing was found, and editor's notes point at a passage for every claim. No uncited praise.
The four levels
The dial is monotonic: each level includes everything below it. It is set per project, so a project you are experimenting with can sit at Draft while your novel stays at Off.
| Level | Name | What it means |
|---|---|---|
| 0 | Off | Nothing leaves this machine. |
| 1 | Ask | Queries, summaries, tag suggestions, and critique, on request. |
| 2 | Suggest | Adds ghost text and rewrite-in-my-voice. |
| 3 | Draft | Adds multi-paragraph drafting proposals. |
What each feature sends
This is the table shown in Settings, in the same order. It lists every feature the app knows about, including some that are still being built; a feature that is not built yet sends nothing. At launch the provider is OpenAI, using your own API key. Nothing else is sent: no telemetry, no other documents, never the whole manuscript.
| Feature | What it sends | Level needed |
|---|---|---|
| Tag suggestions | The current document's text and the existing tag names. | Ask |
| Scene summaries | A scene's text, to generate its summary. | Ask |
| Assistant chat | The active scene's text, any notes pulled in with #Name, and your message. | Ask |
| Story Intelligence | Your question, scene summaries, and the full text of the top matching scenes. | Ask |
| Editor's notes | The scene's text and its scene brief. | Ask |
| Search indexing | Chunks of scene text, to compute embeddings for search. | Ask |
| Ghost text | Up to 500 characters of text before the cursor and 100 after it, plus the scene’s notes and metadata (location, POV, timeline), and the voice profile (stylometric rules and up to 3 exemplar passages). An answer that breaks the voice profile is sent back once, with the same context plus the rule it broke, for a second try. | Suggest |
| Author mode | The active scene's text, referenced notes, the scene brief, and your instruction. | Draft |
Seeing what it costs
Every request records the model, the tokens in and out, and the cost in your project's usage ledger, and the AI tab in Settings shows today's spend and the running total. You can set a daily cap; at a cap of zero the app makes no requests at all, cached answers included. Ghost text has its own brakes: it waits for a pause in typing, needs a minimum of new text since the last request, never asks while a suggestion is showing, and stops after a daily limit.
Where to find it
Open Settings (Ctrl+,, or Cmd+, on macOS), choose the AI tab, and the dial is at the top under “This project”. Your API key, the model tiers, and the usage figures are on the same tab.
In short: every AI output is a proposal, nothing enters the manuscript without your explicit accept, and AI is off until you turn the dial.