Teaching AI, Built With AI

"Teaching AI, built with AI" is not a slogan here — it is a literal description of the working method. I designed a 3-day, on-site generative-AI training programme for a cohort of around 70 engineers: the quantitative analysis of the pre-survey, the curriculum, and the exercise materials themselves. The toolchain was Python, python-pptx, and Claude Code. The thing being taught and the thing doing the building were the same.

The design had to satisfy three constraints at once. First, assume near-zero prior experience with requirements definition, and still get people to feel what upstream thinking is like inside three days. Second, break overconfidence in AI. The failure that recurs in generative-AI training is trusting the output uncritically, and a lecture does not fix it — it has to be built into an experience. Third, leave nobody behind: across a group of ~70 with real skill spread, everyone needs to walk out with something that worked.

The first pillar was the pre-survey analysis. Five indicators were inverted into ability points, combined into a weighted composite score, and cut into three tiers that drove the group structure. What mattered was not the classification but the robustness audit around it. If you hand a client a segmentation, you have to answer "why this line?" on the spot or it cannot be used for a decision. So: a sensitivity check showing 94% of people stayed in the same tier under even weighting; an explicit boundary band of ±0.5 points labelled as genuinely ambiguous rather than silently assigned; and response-quality flags for suspiciously fast or flat-lined submissions. The analysis shipped with a document that separated what the data supports from what it does not. Every step was scripted, so the same input always produces the same output.

The second pillar was exercise design, and the principle was to withhold the answer. In the AI-output evaluation exercise, participants judge the output before being taught any evaluation framework — the reveal comes afterwards. Errors are seeded at three difficulty levels, so every participant, at whatever level, gets to find at least one and know they found it. The interview role-play runs as a card-based format: groups of four, four rounds, 75 minutes. The cards are built to return nothing useful in response to an open question. You can only make progress with closed questions driven by a hypothesis — which turns "come in with a hypothesis" from advice into a structural constraint.

The third pillar was the materials pipeline. A lot of slides had to conform to a brand template, and rebuilding a template from scratch always drifts — a logo, a margin, a fill. So instead of recreating anything, the PPTX gets unpacked, edited as XML with lxml, and packed back up; the template's own decoration is inherited untouched. Verification runs through LibreOffice to PDF, then pdftoppm to images, and the layout breaks and overlapping text get caught by eye. The Japanese-font problem, where CJK glyphs silently fall back to Chinese forms, was pinned down at the rPr level. Fourteen exercise materials came out of this path, sized for 17–18 groups running in parallel.

Three things stayed with me. One: an analysis handed to a client being correct is not enough. It becomes usable for a decision only once you can answer "how do you know?" without pausing. Two: using AI to design the training about using AI gives the instructor a felt sense — not a described one — of where the boundary between machine work and human work actually sits, and that showed up directly in how convincingly it could be taught. Three: do not try to reproduce a template; inherit the XML. It is faster and it is certain. Delivery is September 2026. How well the design held up gets measured by the shift between the pre- and post-survey, and that is a separate piece of writing.