Requirements Definition with Generative AI
Teaching AI, built with AI.
Overview
- Client
- Corporate generative-AI training, contracted through a training provider
- Role
- Technical PM · Instructional design · Data analysis
- Period
- Jun – Sep 2026 (design, build, delivery)
- Scale
- ~70 participants · 3 days · on-site
- Stack
- Python (pandas) · python-pptx / lxml · LibreOffice CLI · Claude Code · GitHub Copilot / Codespaces
The challenge
Three constraints had to hold at the same time.
Assume near-zero prior experience. The programme had to give people the feel of upstream thinking inside three days, in a domain where lectures do not transfer — it has to be run as exercise.
Break overconfidence in AI. The failure that recurs in generative-AI training is trusting the output uncritically, and telling people to be careful does not fix it. It has to be built into an experience where they get caught.
Leave nobody behind. Across a cohort of around 70 there is real skill spread. A programme where the strongest enjoy themselves while the rest sit quietly is a failed programme; everyone needed to walk out with something that worked.
① Pre-survey analysis → group design
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. Handing a client a segmentation means being able to answer "why this line?" on the spot, or it cannot be used for a decision. So: a weighting sensitivity check (94% of participants held the same tier under even weighting), an explicit ±0.5 point boundary band 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 — an analysis you cannot reproduce is one you cannot verify.
② Curriculum — exercises that withhold the answer
The three days map to understand → find the problem → specify it, designed as one line where each day's output becomes the next day's input rather than three self-contained workshops.
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 finds at least one and knows 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 instructor's worked examples use generic material while participants work on their own themes, so nobody can copy the demonstration into their answer.
③ Materials build pipeline
A brand-compliant deck cannot be rebuilt from scratch without drift — a logo, a margin, a fill. So the template is not recreated: the PPTX is unpacked, edited as XML with lxml, and packed back up, inheriting its own decoration untouched.
Output is verified by converting to PDF through the LibreOffice CLI and rasterising with pdftoppm, so overflowing text and broken layout are caught by eye before anyone sees a slide.
The Japanese-font problem — CJK glyphs silently falling back to Chinese forms — was pinned down at the rPr level rather than worked around.
The three-day design
By the numbers
- Participants
- ~70
- Programme length
- 3 days, on-site
- Pre-survey response rate
- 100%
- Tier agreement under even weighting
- 94%
- Exercise materials authored
- 14
- Groups running in parallel
- 17–18
What I took away
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. The sensitivity check and the boundary band were not accuracy work — they were accountability work.
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. That showed up directly in how convincingly it could be taught.
Do not try to reproduce a template; inherit the XML. It is faster and it is certain — and the same instinct generalises well beyond slides.
This case study omits the client name and identifying details out of confidentiality, and is limited to design method and process. Figures are approximate.