The Academy for Climate Jobs: installation checklists generated from the courses that teach them
A training provider for climate jobs had the knowledge in its courses and needed it as checklists technicians can use on site. The generator reads the course, writes the checklist, and refuses to add anything the course did not say.
The repetitive task
The Academy for Climate Jobs trains the people who install the hardware of the energy transition: wallboxes, smart meters and the electrical work around them. The training lives in SCORM course packages, lesson by lesson, with videos. What a technician needs on site is different: a checklist of what to do, in order, in the words the course used, with the relevant video a tap away.
Producing those checklists by hand meant someone reading through each course, extracting the procedural steps, leaving out the theory and the learning goals, keeping the terminology consistent and rebuilding the whole thing when a course changed. For a catalogue that grows with every new course and exists in more than one language, that does not scale, and it is exactly the kind of work where a person copying from a source gets details wrong.
What was built
A web app the academy's team logs into, uploads a SCORM package to (up to 800 MB, videos included), and gets back an editable checklist.
-
Course ingestion
The package is unpacked in the client's Supabase project. Lessons are identified from the course structure, the language is detected automatically, and every video is extracted and stored so it can be embedded in the right step later.
-
Lesson by lesson generation
Each lesson is processed as its own chunk with a strict prompt: extract only what the technician physically does, title each step in the course's own terminology, exclude learning objectives, introductions, theory and quiz content. A merge step then assembles the lesson outputs into one checklist and names it.
-
Editor with live preview
Steps, substeps and headers are editable in place. Videos that could not be matched to a step automatically sit in an unassigned panel for the editor to drag into position. Variants of the same checklist can be kept for different export targets.
-
Export for the field
An interactive HTML checklist with progress tracking that a technician works through on a phone, and a PDF for the clipboard version. Both come from the same source so they never drift.
-
Admin and access
Invite code registration, user management and templates for recurring course types, so the academy runs the tool without a developer in the loop.
Verification built into the generator
The risk with generating safety relevant instructions from a model is obvious: a plausible step that was never in the course. The design answer is not to ask the model to be careful but to make invention impossible to ship.
The rule
Every substep must carry a verbatim quote from the source lesson. No quote, no substep.
The prompt treats a substep without a supporting source quote as invalid and discards it. The sequence and terminology are preserved exactly as the course states them, in the course's language, so a German course produces German imperatives and an English one English ones. The result is a checklist the academy's subject experts can review by checking quotes against lessons rather than re-reading the whole course.
This is the same principle that runs the verified-source content pipeline on this site: the model does the extraction and the drafting, a deterministic rule decides what is allowed to leave the system, and a person signs off on output whose provenance has already been checked.
Content you already have, in a shape you don't?
Courses into checklists, documentation into onboarding, transcripts into briefs: if the source exists and the output has to stay faithful to it, this is a build with a verification layer, not a prompt. Start with an implementation audit or describe the transformation you need.