How the Companion works

Every AI exchange you log lands in one of two buckets. Nobody sorts them by hand: the split follows what happened when you pressed send, and it is about the prompt, not about whether you used the Companion.

• Considered — the prompt was worth sending. Either it was already clear enough that nothing interrupted you, or the Socratic Mirror stepped in and you reworked it before sending. Both count the same, because both mean you thought before you asked.
• Sent anyway — the Mirror flagged the prompt as thin and you sent it unchanged. That is a legitimate choice, and sometimes the right one: verifying a result or unblocking one step does not need a polished prompt. It is only worth noticing when it becomes the habit.

Used an AI tool outside the extension (mobile app, another computer, a desktop agent)? You can log that exchange manually from your Prompts history, so your record stays honest and complete.

The extension reads your prompt locally in your browser- no AI call, no server - and when a prompt looks like it could be sharper, it interrupts with questions before anything is sent. There is no grade and nothing to optimise: the interruption is a nudge to think, and "send anyway" is always available. What it looks for, and when it stays out of the way, is spelled out in How the Mirror decides to step in.

Before a prompt leaves the page, the extension reads it through four lenses - in your browser, never sent anywhere for this. They are not a grade: they are the four things a prompt worth sending tends to have, and the Mirror steps in when too many of them are missing.

• Clarity— a formed request: a sentence that finishes, an action verb ("analyse", "compare", "fix") and enough words to be understood without guessing.
• Context — what the AI cannot invent: who you are or who it is for, the constraints (length, deadline, format) and, above all, material to work on — an excerpt, an outline, code, a draft.
• Iteration — building on the exchange: picking up the previous answer and refining it rather than starting from zero. Real AI use is several turns, and the Mirror gives credit for that.
• Critical thinking — keeping the upper hand: asking for sources or evidence, saying what you will verify, asking for the limits or the counter-arguments of the answer.

What it does not buy

• Stuffing keywords does not count. What is read is distinct words, not repetition.
• Delegating without material stays capped."Do my homework" surrounded by the right vocabulary is still a delegation: without an excerpt, an attempt or a draft, context and critical thinking cannot come from wording alone.
• The scaffolding is not read as yours.When the dialogue compiles your answers into a structured prompt (the part headed "My prior reasoning"), that added structure is not what the Mirror assesses — your own thinking is.

When it steps in

• Only when you open a subject.Follow-ups inside a conversation are never interrupted. The Mirror comes back into a thread only after three thin prompts in a row, and "Leave me alone on this thread" silences it there for good.
• The bar is set by the programme, and it rises gently with your streak of good first drafts: the more you progress, the more the tool steps aside.
• You always decide: improve the prompt, send it anyway, or cancel.

Every dialogue stays with the prompt it belonged to — you can reread each one on your prompt history.

Everything is private by default. Your journal, contract and prompt history belong to you. Your CARE tutor sees a given item only after you explicitly share it (per journal week; recap or full for prompts; on/off for stats), and you can stop sharing anytime. Programme admins only ever see anonymous aggregates — never individual content. Details in the privacy notice.

• The browser extension runs on Chrome for computers (ChatGPT and Claude websites). It is not available on iPhone or iPad — log mobile usage manually from your Prompts history.
• Travelling or on mobility in mainland China: the app is not reachable without a VPN (its hosting providers are blocked there). Plan accordingly.