Methodology

AI workflows: what each one sends, and the preview guarantee

This is the trust contract behind AI Tools. The five workflows are thin clients over the same BYOK engine as AI chat — same keys, same direct-to-provider calls, same cost estimates, same timeouts and retries. Everything below is in addition to that contract.

The non-negotiable rules

  • Nothing is transmitted automatically. Each workflow shows a “what will be sent” preview before Send, with a cost estimate. Pressing Send is the only way data leaves your browser.
  • The preview is not an approximation. The preview pane and the Send button are built by the same code path that assembles the request — you see exactly the text that will be sent, character for character. What you preview is what is sent.
  • Model output is never trusted. Every AI response is rendered through an escape-first Markdown renderer: your text (and the model’s) is HTML-escaped before any Markdown syntax is applied, so a malicious or broken response can’t inject scripts or markup. Parsers extract plain strings only — they never produce HTML.
  • No keys, no endpoints, no network calls exist in the workflow logic itself. It assembles prompts, parses output, and scores quizzes locally; the only network call is the user-approved Send through the shared chat engine.

The five workflows — what each sends

1. Context chat (chat with your study space)

You tick the notes/files to include. The workflow assembles a payload from only your selections: your question plus the selected items’ text, subject to context-window guardrails — at most 20 items, 8,000 characters per item, 24,000 characters total — with per-item truncation noted in the preview. Nothing else from your study space is included. Output token budget: 2,000.

2. Flashcards from notes

Sends your chosen note/file text plus an instruction to return question/answer pairs. The returned pairs are parsed locally into cards; only pairs with both a front and a back are kept. You review and edit the deck before saving it.

3. Quiz from notes

Sends your chosen note/file text plus an instruction to return multiple-choice and short-answer questions (defaults: 5 MCQ + 3 short). Scoring runs locally: MCQ is auto-scored against the parsed answers; short answers are shown for self-check and are not machine-graded. The attempt (score, answers, timestamp) is saved locally — see storage below.

4. Planner from syllabus tracking

Sends a summary of your tracker state — subject names, coverage percentages, weakest topics, days remaining — plus your daily minutes and horizon. It does not send your full notes or files. The model returns a draft study schedule, which is validated locally (subjects must exist, minutes must be positive) and converted into a real timetable you can then adjust and save in the timetable generator.

5. Notes assistant (summarize / expand / simplify / key points)

Sends the note text you selected plus the action instruction. The result is shown as a preview that never mutates your note — you explicitly choose to replace, append, or discard. Your original note is untouched until you decide.

Where generated content is stored

  • Flashcard decks and quiz attempts live in your browser’s IndexedDB, inside the existing kv store (flashdecks:v1, quizattempts:v1) — the same storage as your notes, under the same Study Space data rules. No new database version was needed.
  • Decks are capped at 500 cards; quiz attempts keep the most recent 200. Both are included in backup export/import.
  • Planner drafts are not stored until you save them as a timetable — a draft you abandon leaves no trace.
  • Context-chat messages are part of AI chat history (see AI chat methodology) — local only.

Output token budgets

Each workflow caps the model’s response length (worst-case cost basis for the pre-send estimate): context chat 2,000, flashcards 2,000, quiz 3,000, planner 2,500, notes assistant 1,500 tokens. Smaller responses cost less — the estimate always shows the ceiling.

← Back to methodology overview