How Henri calls a model

The teaching system prompt is the stable part of every lesson call. Its sections cover the message structure, a curriculum that stays fixed once teaching starts, the teaching philosophy, a short welcome, one chapter at a time, what to do when the learner is confused, response length, tone, and a usage guide for each tool. On Claude that prompt is the cached prefix, so later turns in the same lesson reuse it. The tree prompt asks for 12 to 40 chapters, grouped under a few top-level branches. The questionnaire prompt stops once it knows the learner’s motivation, starting point, goal and preferred approach, usually after two or three rounds.

The tool definitions match the evaluation harness used to compare teaching models. generate_quiz returns a true-or-false or a choice question, and is_final marks the question that can complete the chapter. evaluate_understanding either advances with reveal_next or switches with switch_to. generate_image asks for one picture of the idea being taught. The evaluation request also defines show_html_widget. The iOS lesson omits that tool, because the phone has no renderer for it. The app checks the arguments against that tool’s schema before it stores the result. A quiz that the learner has not answered is removed from the history before the next ordinary message, because the provider requires every tool call to be followed by its result.

Finished chapters are folded before the model sees them. The saved conversation stays complete. The copy sent to the model replaces each finished chapter with one learning record: three to eight notes about what the learner understood. If that record was never written, the model sees a short completed-chapter marker. A record that is already in that form is left as it is. The opening turn that carries the questionnaire and the curriculum stays. On the adaptive lesson, the chapter being taught keeps every turn. On the earlier lesson path, that chapter keeps the latest 30 turns, and the window is lengthened when it would otherwise start on a tool result whose tool call had fallen outside it. Tool calls that lost their results are removed, so the provider still receives a valid sequence.

Uploaded pages are handled on the copy, too. PDF and file blocks are left out of the compressed teaching history. In the evaluation harness, the original upload becomes a placeholder once the curriculum exists. When the chapter changes, the previous chapter’s pages come off and only the current chapter’s pages are attached for that call. That change stays in memory, so the cached prompt prefix is unchanged.

A normal lesson uses one chain. Claude Haiku on Bedrock is first. Claude Haiku through OpenRouter is next, and that step is omitted when its key is missing. Official DeepSeek V4 Pro is last, with thinking turned off, because that is the setting under which the teaching evaluation selected it. Each attempt has its own timeout, and the whole chain has a deadline. The first attempt always runs. A later attempt runs only when enough of that deadline remains.

Each attempt receives the schema in the way that provider can accept: the native JSON schema, JSON mode with the full schema added to the prompt, or the full schema in the prompt alone. Required fields, types and allowed values travel with the schema on the prompt path.

Two other jobs use a shorter chain of their own. Adaptive teaching starts on official DeepSeek with high reasoning, then falls back to Claude Haiku on OpenRouter. Changing the learning plan starts on GPT-5.6 Terra at medium reasoning through OpenAI, then the same model on Azure, and then the normal lesson chain. A missing key skips that step. The saved call records the provider, the model and whether a later step answered.