Lens
An analysis engine I built that audits a training program for instructional-design quality.
- Quality checks in the catalogue
- 78
- Findings on one course series
- 130
- Audits run on real programs
- 11
Node.js, Claude Code, MiniLM embeddings, docx
When I review a training program, I start by deciding which questions matter for this audience and this intent, and then I read the materials for those. A checklist would run every rule on every document. I built Lens to audit a program the way I would, and to write down what it found in a report I can hand over.
What I built
Lens reads the finished materials (PowerPoint, Word, PDF, Excel) with a short brief, and writes a findings report in Word.
- Intake and a gate
- Lens drafts the brief from the materials and then stops at one question: what is this program meant to change, and for whom? It audits nothing until I have answered.
- It chooses its lenses
- From the brief, Lens picks the audit modules and overlays that fit the subject and the format, and it writes down why it picked them.
- 78 checks
- The catalogue holds 78 named quality problems. Each finding comes with a severity, the place the evidence sits and a correction.
- Three reviewers push back
- Before the report is written, a cohort lead, a subject-matter expert and a sponsor each argue against the findings.
- It remembers
- A finding I dismiss stays dismissed, and when Lens audits the same program a second time it marks each finding as new, resolved or recurring.
What the first real audits changed
The third real audit was a technical course on industrial bearings. Lens returned more than ten findings, and all of them were one problem: the assessment only tested recall. I marked up the report by hand, and three rules came out of my notes. Lens now merges findings that share one root cause into one finding. It keeps its recommendations inside the course unless I ask for a follow-up plan. Its recommendations address the learners taking the course and nobody else.
What it delivers
Lens has run eleven audits on two real programs so far. Across the three modules of the bearings series it logged 130 findings, and it wrote a short synthesis of the patterns that ran through all three. Those audits set the brief for the revamp of the series, and the idea of turning the series into a game came out of the same work. That game became the flagship behind the Learning Game Engine.
