coursekit

Español

AI-assisted e-learning course production: reference material, instructional design, writing, review, media, assembly and SCORM delivery, with human sign-offs at the control points. It works with Claude Code, opencode and Codex, interchangeably.

Package slxd-coursekit, command coursekit. By Studio LXD.

Early development (0.5.0). Published on PyPI as slxd-coursekit.

What it does

coursekit turns a folder into a course production project and drives the work through a fixed process in which AI agents do the drafting and people keep the decisions:

1. Brief — you drop the reference material (documents, links, notes); coursekit converts it for the agents.
2. Instructional design — an agent proposes units, sections, objectives and hours; you sign it off.
3. Writing — an agent writes every unit in a fixed Markdown format; coursekit verify checks it against the production rules.
4. Review — a second agent (another tool or model) reviews each unit; you sign off each unit.
5. Media — images, graphics, video, audio and interactive demos are planned and produced with the tools you have.
6. Assembly — from the approved Markdown, never the other way round, with one of two backends: the content is loaded into the authoring platform (slxd creator), or coursekit itself builds a SCORM package per unit (the html backend, no platform in between). You choose per project or per course.
7. Delivery — SCORM packages (exported from the platform, or built by coursekit) are recorded and published to a shared folder with a tracking spreadsheet.

Only a person can sign (coursekit approve); agents are forbidden to run it.

Status

Early development (0.3). What is solid and what is still to prove:

• Used end to end: the process from brief to SCORM delivery, with both backends (slxd creator and the html backend), in Spanish and English; the decision-based approach with simulations; the local instructional design; the handoff mode.
• New and less proven: migrating an existing course from eXeLearning packages (coursekit import, /update-unit, videos from official channels). It converts real packages without losing text, but the whole flow with real agents is still being piloted.
• New: videos as data (coursekit video): a montage of scenes in the theme of the course, from a library of atoms, molecules and scenes that the project and each course extend, each scene lasting what its voice lasts; and its timeline for an editor (coursekit video export: DaVinci Resolve, Final Cut, Premiere).
• New: checks of the audios and videos produced (coursekit media check: loudness, captions too fast to read, what is on screen and not said), their accessible transcript and track of descriptions (coursekit media transcript), a contact sheet to look at a video without playing it (coursekit media sheet), and zooms of a screen recording that follow the narration (the Screencast component).
• New: the Gemini TTS voice (provider gemini, with a style instruction and the model fixed in each course), at half the price: in the flex tier by default, or in batch for a whole course (coursekit tts-batch: queued with --batch, collected hours later, by the handoff too). Tried against the real service; the batch has not yet carried a whole course. The long scripts of Azure, Google and Gemini are now joined as uncompressed audio and encoded once.
• New: the brand in the media. coursekit brand reads the brand material left in theme/branding/ (a brand book, a Storybook read whole, the website of the company observed in a browser) and takes out its fonts, logos and tokens; the theme agent writes from it the style guide of the media, which a person signs (coursekit approve media-style), and proposes the colour of the brand for each role the platform lacks (project.yaml › theme.roles, coursekit theme roles). The fonts of the brand go inside the videos and the SVGs exported as PNG. Tried with a real Storybook and website; the agent step has not yet carried a whole project.
• New: illustration sources of the brand. The Magnific page of an author whose illustrations fit the brand goes in theme/branding/links.md; coursekit brand keeps previews of their work as style references, and coursekit media stock finds among their illustrations the one a concept needs, takes it as SVG and recolours it to the ink and paper of the brand (media › stock), with its origin and licence in the manifest. Needs MAGNIFIC_API_KEY; not yet used on a whole course.
• New: where the tests go (delivery › assessment): inside each unit, as before, or apart in one package of their own after the units (placement: separate); and their questions as lessons of the packages or as a question bank the LMS imports (format: aiken | gift | moodle-xml | qti: with the creator backend creator exports it from a content of its own, with html coursekit delivery questions writes it), with verify refusing the question types the format cannot take. Tried with the html backend and the files coursekit writes in the tests; not yet assembled apart in creator nor imported in a real LMS.
• New: inclusive language (rules › inclusive_language: none, neutral by default, split, slash): the design, the writing and the review follow it, and verify warns about the generic masculine and the slashes with a lexicon per language.
• New: the same review Excel for a design made in slxd as for a local one (coursekit design CODE writes it from matrix.json and validation.json): competencies, objectives with where they are developed, practised and assessed, hours and minimum words per section, grading of the tests, common errors, and the validation with the expected warnings apart. The Excel slxd exports is kept as a copy, in the language of the person.
• New: official sources through a web search API (coursekit research: Exa, Perplexity, Brave or Tavily, chosen in the .env). The design looks for the reference documentation with it and the writer for the source of a fact the brief does not have; it only proposes addresses, in the language of the course when the site has a translation, and they enter the course through brief/links.md. Without a provider the agents use their own web search. coursekit brief now opens in the browser of the machine the official pages that refuse a direct download (EUR-Lex, for one). Tried against the four real services; not yet used by the agents on a whole course.
• New: lessons of a whole pilot course. Each course keeps a record of what people decided and of the interface labels they checked (decisions.md), which writing and review respect, and the reviewer checks every pending item against the current state before repeating it. verify warns of an activity not written with the directive of its type, of a test question that repeats a practice situation and of a simulation error with no consequence; the writing rules cover sorting questions with one order, feedback and the current interface only. The media manifest is locked while it is written (agents in parallel), media render retries the browser, brief converts linked PDFs, the style guide of the media quotes every line of notes.md (and approve media-style warns of those left out), speakers can go without a portrait or with a brand avatar, and several scripts are signed at once. Not yet tried on a whole course.
• In progress: captures of real applications that coursekit makes itself (coursekit capture): profiles of the web applications, the permission to capture, and covering the personal data of a recording a person provides (OCR, reviewed and approved by a person) are in place, and so is an agent driving a web application in the browser with the personal data replaced on screen (MCP server), with its video recorded by repeating the steps (coursekit capture replay), the course itself (its html package served locally or its review link in creator, with no login), and a window of a desktop program (covered on the pixels): on macOS; on Windows and Linux with X11 it is written but not yet tested on real machines.
• Known gaps of the creator platform that the pipeline works around: Known limits of the creator backend.
• Not yet: reading eXeLearning 4 packages (.elpx) or HTML exports of a site, publishing the components of the html backend as a package of their own, more languages than Spanish and English.

Install

You need uv; it brings the right Python (3.12 or 3.13). To work on courses you also need git, an AI tool (Claude Code, opencode or Codex) and the URL of your SLXD Creator MCP server; see Getting started.

uv tool install slxd-coursekit
coursekit --version

To install the development version straight from the repository instead: uv tool install git+https://github.com/studiolxd/coursekit.

To try it from a clone, with changes taking effect immediately:

uv tool install --editable /path/to/coursekit

The real captures of applications (coursekit capture) need the optional capture extra: uv tool install --force "slxd-coursekit[capture]" (from a clone: uv tool install --force --editable "/path/to/coursekit[capture]").

To update an installed copy: uv tool upgrade slxd-coursekit. Nothing else: the first command in each project regenerates its managed files and agent files. A specific version: uv tool install slxd-coursekit==0.2.1.

Quick start

coursekit init my-courses        # a short wizard: language first, then everything else
cd my-courses
coursekit doctor                 # what is installed and configured here

Then open your AI tool in that folder and run:

/new-course "Strong passwords" 2

or, from the terminal, without opening anything yourself:

coursekit run new-course "Strong passwords" 2

coursekit status always tells you where each course is and what the next step is.

Languages

The tool speaks Spanish and English: the wizard asks the language first and everything after (messages, help, spreadsheets, templates for people) follows it. Courses can be written in either language; the content format adapts to it.

Documentation

Full documentation, in English and Spanish, is in docs/ (versión en español):

Contributing

See CONTRIBUTING.md and AGENTS.md (for coding agents working on the package).

License

MIT — see LICENSE.