Imported from silkansraivis3-tech/course-factory (course-factory-plugin/skills/course-factory/SKILL.md). Install upstream with npx skills add silkansraivis3-tech/course-factory --skill course-factory. Copyright stays with the author.

Course factory — an accredited programme becomes two tablets

v1 (2026-09-07) · Maintained by Raivis · Personal — installed for Raivis only, not org-published

The acceptance test. Someone who has never seen the course opens the instructor terminal, types their ID, picks the module, presses Start, and presses Next to the end — and a real accredited class gets taught, in the minutes the approved programme allows, with the trainees' tablets carrying every task the screens announce.

This skill is the entry point for that delivery target. It decides mode, runs the gate, and routes. It does not itself write screens, tasks, or content.

Mode first, then read one lane

Never load a lane the mode does not need. hours/ answers "how long may this be"; tablet/ answers "where does it live and how is it wired"; coverage/ answers "is everything the programme and the model course ask for actually taught". They never run together.

coverage/ exists because the other two cannot see its failure. A course can have exact hours and perfect wiring and still not teach an outcome. On GAS BASIC the topic-level record was green at 64 of 64 and had never once been checked against the model course's own 266 itemised outcomes — where one real gap was waiting.

Deliberately outside this bundle — point at these, never copy them

If a question belongs to a row above, say so and point. Do not re-derive it here.

The brief — what the operator types, and what you must not ask

The operator types free-form. Extract these; ask only for what is genuinely absent, batched 2–4 at a time:

1. Which accredited programme governs — the approved document, by name or path. Without it there is no hours law and no ILOs, so this is the one blocking question.
2. The knowledge base — the docling-extracted sources for the subject.
3. What is different from GAS BASIC — simulator? live-fire facility? group size? days?

Never ask: the module split, the screen count, the task codes, the pass marks, or the file layout. Those are derived, and deriving them is this skill's job. Never web-search during intake — not in the brief, programme or knowledge base means UNKNOWN.

The four laws

These are not preferences. Each one is here because breaking it cost real rework on GAS BASIC.

1 · The approved programme is law, and minutes are the law's units

hours/knowledge/hours-rules.json is the authority — read it, do not restate it. The short form:

• An academic hour is whatever the programme says it is (GAS BASIC: 40 min). Never assume 60.
• Every module's built teaching time must equal its allocated minutes exactly.
• Anything that will not fit becomes self-study in the handout. The handout is the overflow valve; class time is fixed. (Owner's ruling, 2026-09-07.)
• Run hours/scripts/check_hours.py before believing any plan. It exits non-zero when the sum is wrong, and it is the gate.

What this law is for. GAS BASIC was built to 2175 minutes against an accredited 1640 — a 33 % overrun discovered only after all eight modules existed, because nothing counted minutes until someone thought to. The contract that orders the whole course factory (COURSE_START.json) has no hours field at all. This lane adds one.

2 · Main ILOs are untouchable

Main ILOs and the approved study programme are copied verbatim and never edited, renumbered or "improved". Sub-ILOs may be re-expressed to make a topic digital or practical — that is the point of the exercise — but:

• a re-expressed Sub-ILO is marked PROVISIONAL until the owner ratifies it, and
• it is never renumbered in anything a trainee sees.

IMO model courses are guidance; the approved programme governs. Where they disagree, the programme wins and the model course is cited as support, never as authority.

3 · 80/20 practical, or active learning instead

Aim for 80 % practical / 20 % theory. On an accredited programme this is often arithmetically impossible — GAS BASIC allocates 8 practical hours of 43, i.e. 19 %, fixed by the programme's own table. When the split cannot be reached:

the theory hours are delivered as active learning — a tablet task inside the theory block, not a longer lecture.

That is the fallback, and it is mandatory, not optional. A theory block with no trainee activity in it is a defect in this system. Record the achieved ratio in the companion file with the reason.

4 · Nothing is useless until you have grepped for what depends on it

Before removing any control, label, field or file: grep its id, class, label and filename across both terminals, the course tree and every generated script. On GAS BASIC, three of four "useless" findings were load-bearing — timer buttons the instructor run script tells the instructor to press, print buttons on hand-out documents, and hrefless <span id="gb-home"> elements that are how a page declares it has no way back. Unused today is not useless: data a planned system will read is not dead.

The gate — approved before anything is built

Changing a table row is free. Changing eight built modules is not. Present this and stop:

Nothing is built until the operator approves that table. The module split and the minutes are the expensive things to get wrong.

Build order

knowledge/build-order.json is the authority. Read it; it names the owning skill for each step and what each step must not start without.

The shape, for orientation only: programme → hours → module split → gate → ILO map → per-module screen inventory → screens → tasks → handout → assessment → terminals → verify.

Two orderings that are not negotiable, both learned by getting them wrong:

• Screen inventory before any screen exists. course-module-ux owns that gate.
• Task pages before the deck announces them. A screen that announces T3 when the tablet has no T3 is the worst failure in the room, and it is only caught by the three-way cross-check in tablet/.

Honesty markers

One set, identical to the rest of the family (novikontas-course-start owns the definitions):

Never invent regulatory content, hours, pass marks or citations. Every marker must appear in the companion file — a marker used and never reported is the failure this rule exists to stop.

Verify before reporting done

Run these, read the output, then report. A green exit is not proof the work happened — inspect the line for the thing you changed.

1. hours/scripts/check_hours.py — minutes match the programme
2. coverage/scripts/check_syllabus_coverage.py — every itemised outcome of the model course is taught. Not the same question as the topic record, and it found the only real content gap in 266 outcomes on a course whose topic record was already green.
3. tablet/scripts/verify_course.py — structure, counts, check→hand-off order
4. tablet/scripts/verify_links.py — every link resolves in the merged asset root
5. tablet/scripts/audit_navigation.py — every page is reachable and has a way back. Not the same question as (4): a section whose own index nothing links is invisible in the built app however good its links are. That was 28 pages on GAS BASIC — every safety brief, rotation plan and practical write-up — with verify_links.py green.
6. tablet/scripts/crosscheck_tasks.py — deck ↔ run script ↔ tablet manifest agree
7. course-module-ux's measurement lane for the screens themselves

(2) and (5) exist because (1), (3), (4) and (6) all passed on a course that was missing an accredited outcome and shipping a dead section. Every check here was added after something got through the others.

Then write templates/factory-notes.md into the course folder: what was produced, every marker with its source, the achieved ratio, what was moved to the handout and why, and a final section of feedback on this skill. That last section is the only channel through which real use reaches the skill — it is written by this skill during the run, never left for a human.