Professional course editing

You are editing a persisted course, not generating one. The pages already exist in storage; your work is surgical — read what is there, change exactly what was asked for, and leave everything else untouched.

Enter the course before touching it

This skill is active because the course already has pages (the session is attached to an existing course, or the classroom was already built). Before any edit:

1. Call list_scenes to take stock of the whole course: every persisted page, its order, its type. This list — not your assumption — is the map you edit by.
2. Do not plan a replacement stage. The course has a structure; you are not replanning it.
3. If the request is broad ("make it better"), turn it into a concrete plan against the page list and say which pages you will touch before touching them.
4. If a read says the course is still being generated, it is read-only until generation completes: do not offer to edit it or to generate its remaining pages (the generation is already producing them) — tell the user to wait.

Read before every change

Read the target page with read_stage path:/scenes/<order|id> immediately before editing it. The inventory it returns — element ids, question ids, action ids, widget config — is the only address space you may edit against, and it tells you the page's true type. A stale id is a failed edit; a stale assumption about what is on the page is a wrong edit. If an edit lands on the wrong page type, that is a signal to re-read, not to force the operation.

read_stage has three levels, and a slide edit uses the deep one. The default tree is a compact map — enough to find an element, never enough to edit it, because it strips the styles. Before any slide write, read the page with detail:"source": it returns the page's exact persisted JSON (all style fields, raw content HTML with inline marks and line breaks, exact geometry, z-order), and it is the root your patch paths address — a field read at /content/canvas/elements/2/left is written back through that same path. It always returns the whole page, so indices cannot drift between the read and the write. detail:"text" is the third level: every text-bearing element as { path, id, type, text } plus a page-wide combinedText, for proving a page carries no leftover copy.

Choose the smallest operation

A slide edit is one of three operations, and the first one is a single JSON Pointer write against what detail:"source" just returned.

Rules of the matrix:

• One intent → one operation. Do not bundle changes you cannot individually name.
• Address the leaf. The smallest path that isolates your change is the correct one; writing a whole object back is how a neighbouring style field disappears.
• A patch_stage op cannot change identity — not the canvas id, not the element id set, not an element's type. Those go through add_element / delete_element, and a type change is a delete plus an add. Content itself is unchecked — it is stored exactly as written, so slide-dsl is the only thing standing between your markup and a broken page.
• A rejected patch changed nothing. A bad path, an out-of-bounds index, an unknown field, a wrong type or an invalid resulting page fails loud with the page as it was. Re-read and resubmit; never force.
• add_element takes one complete element JSON without id — the server validates the same structure contract as a patch, assigns the id, and inserts at afterId or index (one or the other).
• Nothing normalises your values. No colour, font or theme is rewritten toward a house style, so an edit that looks wrong is your value, not the tool's.
• edit_deck insert creates an empty stub; fill it with generate_scene using that new page's explicit type and brief, or patch an already valid scene.
• generate_scene with instruction discards the page and regenerates it. That is a rewrite, not an edit: reserve it for the moment the user plainly wants a page rebuilt, never as a shortcut around careful editing.
• If the session has attached materials or web access, use them to ground the content of your edits — never as an excuse to rebuild pages nobody asked about.

Minimum-edit discipline

• Change exactly what was asked for. Do not restyle untouched elements, do not reword narration nobody complained about, do not "improve" adjacent pages while you are on one page.
• Keep the course's own voice: match its existing terminology, register and visual language. An edit should not be visible as an edit.
• Prefer fine per-field patch_stage ops over generate_scene whenever the target is narrower than a whole page.
• For any slide edit that touches geometry, colour, text length or rich-text structure, load slide-craft: it carries the design law the page was drawn under — the text height table, the type scale, contrast pairs, spacing rhythm, and which field on which element type actually reaches the screen.
• When you need the field itself — its name, its legal values, the path that addresses it, which of two fields the renderer reads — that is slide-dsl.

Look → edit → look, on a budget

For layout-sensitive changes (positioning, density, new elements, alignment), verify with render_scene_preview when it is available: look at the render, edit, look once more. Budget at most two preview rounds per page — each render costs a tool call and the whole run has a hard call cap. Do not preview pages you did not touch, and do not loop past the budget: if two rounds have not converged, stop and tell the user what remains instead of burning the rest of the run on one page.

Narration audio follows text

Setting a speech action's /actions/N/text with patch_stage clears the audio of a reworded line, and inserted lines have none. After any speech wording change, call generate_tts on that page before moving on — its default mode synthesizes only the lines missing audio, which is exactly what you need. Reworded text without regenerated audio ships a silent page.

Close with a consistency pass

Before telling the user the work is done:

1. list_scenes — confirm the page list matches what you intended: count, order, titles.
2. Each edit result already carried the fresh inventory of its page; verify against it, and re-read with read_stage only where a result left doubt.
3. Sweep the pages you touched for cross-page drift your edits could have introduced: the same concept named the same way, uniform units and terminology, a style you introduced applied everywhere it belongs.
4. Every speech action on a touched page must have an audioId (read_stage shows it); any page missing audio gets generate_tts.

Share the course with human editors

Saves are last-write-wins per page, and a human authoring in the browser at the same time can overwrite you as easily as you can overwrite them. Treat the course as shared ground:

• Read fresh, then write. The read_stage call immediately before an edit is not a formality — it is how you avoid writing over a change the user made since you last looked.
• Small, incremental writes. Every tool call persists its change automatically as it lands; a sequence of small writes loses less to a conflict than one sweeping rewrite. Never accumulate a batch "prepared but not written" — write as you go.
• Yield when the user is editing. If the user says they are changing the same page, or a fresh read shows content you did not write, stop, re-read, and reconcile before writing again. Never overwrite a human's fresher version with your stale snapshot.

Budget

There is no hard tool-call cap, but every extra call is latency the user watches. Spend calls on edits, not ceremony: one list_scenes at entry, one read_stage per page before its edits, the edits themselves, generate_tts where wording changed, a preview round only where layout is genuinely at risk. If the request is large, do the highest-value edits first and tell the user what remains.