Skip to content
Goodmargin

How AI reads your writing

Goodmargin uses an AI model to read drafts, suggest which of your own stories might fit a prompt, and notice repeated material across your application. This page describes what the code actually sends and keeps. Nothing is sent until you start one of these features, and you confirm what’s included the first time.

What each feature sends

Essay feedback

  • The text of the draft, frozen as a version that never changes afterward.
  • The prompt you attached (if any), the essay type, and the word limit you entered.
  • Not sent: your email, other essays, private notes, My Story, colleges, or activities.

Story suggestions (brainstorming)

  • The prompt you enter, plus only the story cards and activities you tick for that run. Items you’ve excluded from AI analysis can’t be ticked.
  • Your profile notes, only if you turn on “Include in AI analysis” for them.
  • If you’ve turned on application checks: the titles and opening passages of essays you’ve included, so the suggestions can mention overlap.

Application check

  • Runs only if you turn it on. Sends the latest saved version of each essay you’ve included and the activities you’ve included.
  • You can exclude any essay, story card, or activity, or turn the feature off, in Settings.

Where it goes

Requests leave Goodmargin’s server for OpenRouter, which routes them to the model’s provider. The pinned model is anthropic/claude-sonnet-5. We don’t use a moving “latest” model, so feedback stays consistent. Every request asks OpenRouter to use only providers that don’t collect prompts for training (data_collection: deny). Zero-data-retention routing is not currently required, so a provider may keep request data for a limited time under its own policy, for example for abuse monitoring. See OpenRouter’s provider data practices.

Your browser never talks to the AI service directly, and API keys never reach your browser.

What Goodmargin keeps

  • The result, stored with the exact version, prompt, model, and rubric version it used. Reopening feedback shows the stored result. It doesn’t run again or change wording.
  • A usage record for each request: time, feature, model, success or error, duration, token counts, and cost. It never contains your text. Server logs follow the same rule.
  • Goodmargin doesn’t use your writing to train models and doesn’t run third-party analytics.

What the AI is told

The reviewer is instructed to point at your exact words, ask questions, and suggest revision goals. It is told not to write sentences for you, predict admission, or invent anything about your life. Before any feedback is shown, the server checks that every quoted passage appears exactly in your draft and withholds anything that doesn’t, or that looks like replacement wording. AI feedback can still be wrong; you can report any comment from the feedback page.

Your controls

In Settings you can turn AI analysis off, turn application checks off, export everything, and delete your account. In My Story and each essay’s details you can exclude individual items from analysis.