AI weekly plan

The plan is your typical week: for each day, the sessions — subject, duration and a concrete activity. You can create it by hand, start from the example, or generate it with AI.

Week: the timeline of phases up to the exams and each day's plan, with completed sessions ticked off.

What the AI receives

• Subjects: name, code, type, load, ECTS (European credits, a measure of workload), exam date.
• The suggested share of time and measured mastery of each one (see time allocation).
• Weak foundation: prerequisites not yet taken, failed, or passed with a low grade.
• Start and exam dates, hours available per day and your notes.

The rules it must follow

The system prompt requires the AI to apply:

1. Retrieval practice in every session — exercises, questions without notes, explaining from memory. Never rereading, highlighting or summarising as the main activity.
2. Spacing — medium and heavy subjects on at least 2 non-consecutive days, with short reviews 1–3 days later.
3. Interleaving — mixed exercises; a main session plus a review of another subject on the same day.
4. One weekly cumulative retrieval session.
5. Blocks with breaks, never exceeding each day's hours.
6. Sleep and load — if the hours aren't enough, cut light subjects and languages first, never the spacing of the heavy ones.
7. Languages in short, frequent sessions.
8. Time allocation according to the suggested share; the nearest exam takes priority.
9. Weak foundation → short retrieval sessions on the prerequisite in the first few weeks.
10. First exposure in blocks; then mixed practice.
11. Phases up to the exam, ending with mock exams under exam conditions.

The response is structured JSON (with a schema), cleaned and validated before it appears: days 0–6, only subjects that exist, dates in the right format.

Scientific check

Even so, the AI can make mistakes. The editor runs fixed rules on any plan:

Rule	Warns when
Load	A day goes more than 15% over the available hours
Coverage	A subject has no sessions
Spacing	A medium or heavy subject appears on only one day (reviews tagged with its code count)
Retrieval	There is no test, recall, mock exam or cumulative review session
Passive techniques	There are sessions based on rereading, highlighting or summarising
Languages	Fewer than 3 sessions a week

Time allocation

A heuristic, not learning science

There is no evidence linking time per subject to ECTS. This rule decides how much time; the principles decide how. The parameters are adjustable and should be checked against your closed-book tests.

priority(s) = ECTS(s) × deficit(s) × [2 if exam in the next 7 days] × [1.25 if weak foundation]
deficit(s)  = 1 − mastery(s)          — from unaided tests only; with no test, 1
              + 0.2 if overconfident
share(s)    = max(10%, priority / Σ priorities), renormalised

• Before the first test, the split is proportional to ECTS (without ECTS, it uses the load).
• No subject drops below 10% — spacing never lets it go forgotten for weeks.
• Subjects whose exam is already done leave the allocation.

Next to each share: why (mastery, exam, weak foundation) and how much it has in the current plan.

Providers

AI_PROVIDER	How it's called
gemini	REST generateContent with responseSchema
anthropic	Official @anthropic-ai/sdk SDK, structured output (output_config.format) and a server-side fallback if the model refuses
openai	chat/completions with response_format: json_schema — works with OpenAI, OpenRouter, Groq and others