Every teacher deserves an instructional coach; almost nobody gets one. Watching one lecture and writing careful feedback costs a coach the better part of an afternoon, so most recorded teaching is never reviewed at all. The AI report on this page cost about 19 cents to produce (the receipts are under “How it works” at the bottom).
So I pointed an AI coach at one of my own graduate lectures — a humbling thing to publish, but it seemed only fair to grade myself first:
A live class session, recorded as it happened.
Mux writes a timestamped transcript and grabs still images of the slides.
Claude scores the teaching and the slides, and has to back every score with a direct quote.
Every quote jumps the video to that exact second, so you can check the feedback yourself.
From a graduate policy lecture, University of Chicago Harris School, Spring 2022
You teach data literacy with real skill and warmth. The strongest thread is how you use the audience's own experience as college graduates to expose biased intuition, then walk through charts progressively — all students, four- versus six-year windows, then racial subgroups — so learners see how analytic choices change the story. Your slides support this well: they're clean, headlines state the takeaway, callouts sit right on the chart with leader lines, and you reveal annotations one layer at a time. You also tie the numbers to authentic advising decisions, which gives the data purpose. The main growth area is that this stays a receptive walkthrough — learners watch rather than practice. The opening chart was too dense (students voiced confusion about the axes), and your closing reflection slides are text-only. Naming an explicit task up front and adding a quick prediction prompt would lift engagement.
M. David Merrill spent decades studying what makes instruction actually work and distilled it into five principles (First Principles of Instruction, 2002). The AI awards a star level for each one based on how much of the rubric it can prove — click any quote to jump the video to that moment.
Is the instruction presented in the context of real-world problems?
You anchor the material in a genuine real-world concern — college-completion gaps and how staff advise students — which gives the data purpose. However, you present it as information to interpret rather than framing a concrete task learners will be able to perform, and there's no deliberate progression of problems for students to work through. The problem context is real but stays in the background of a data walkthrough.
Does the instruction attempt to activate relevant prior knowledge or experience?
You repeatedly activate learners' own lived experience as college graduates, using it as the hook for the core insight about biased intuition. This is a strong, deliberate move that connects new data to what the audience already knows. It falls short of gold because the activation is rhetorical rather than structured recall or a task drawing on prior knowledge.
Does the instruction demonstrate (show) what is to be learned rather than merely tell?
You show rather than tell throughout, walking learners through charts and progressively revealing how grouping, time windows, and subgroups change the story. You use multiple representations (all students, Black and Latinx combined, Black alone) and explicit comparisons, and you make the data-visualization craft itself visible. It misses gold because the visual reference is hard to follow in the transcript and some explanations left students confused about what the axes represented.
Do learners have an opportunity to practice and apply the new knowledge or skill?
Learners do not practice interpreting data or making choices themselves in this excerpt; the activity is receptive. Their contributions are clarifying questions rather than attempts to apply a skill with feedback. This is largely a function of the lecture format, though you could have posed a quick prediction or interpretation prompt to invite application within it.
Are learners encouraged to integrate (transfer) the new knowledge into their own world?
You gesture toward transfer by tying the analysis to real advising decisions and the tension between grad rates and student values like choosing an HBCU. Learners are invited to reconsider their own biased intuitions, which is a form of personal integration. However, there is no opportunity in this excerpt for learners to publicly demonstrate, defend, or create their own use of the skill.
Richard Mayer has run decades of experiments on how people learn from words and pictures; his principles are the standard evidence-based checklist for slide design. A second AI review looks at eight still frames from the video and checks the slides against five of those principles.
Slides exclude extraneous words, pictures, and decoration that don't support the point.
Slides are clean and free of decorative clutter — the title slide at 00:26 uses only a faint watermark logo, and the chart slides carry nothing beyond the graphic, its labels, and the institutional footer. No stock photos or distracting embellishments compete for attention.
Key material is visually highlighted — the eye is guided to what matters on each slide.
Each slide headline states the takeaway in bold (e.g., 'This attainment is skewed by college' at 02:23 and 'high graduation rate seats is much lower for minority students' at 06:19), guiding the eye to the point. The two-tone shading of the area chart also cues the split between above- and below-50% completion.
Labels and annotations sit next to the part of the chart or graphic they describe.
Callout boxes sit directly on the chart with leader lines pointing to the exact region they describe — at 04:21 the '24% of seats below 50%' and '22% above 80%' labels connect to their portions of the curve. Axis titles also sit adjacent to their axes.
Complex content is broken into digestible parts rather than one dense wall.
You build the same chart progressively across 02:23, 04:21, and 06:19, adding one annotation layer at a time rather than showing all callouts at once. This staged reveal breaks a dense graphic into digestible steps.
Ideas are carried by graphics plus narration, not narrated text dumps.
The closing 'A few data reflections' slides at 12:11 and 14:09 are pure bulleted text with nested sub-bullets and no accompanying graphic. This portion relies on a text dump rather than pairing visuals with your narration.
All three calls run on a single model — Claude Opus 4.8, Anthropic's current top-tier general model (the expensive one). The receipts: the two input counts below are measured with Anthropic's token counter; the outputs are estimates from the stored report, so treat those as roughly right.
| Call | Input tokens | Output tokens |
|---|---|---|
| 1 · Rating the teaching | 6,799 | ~2,500 |
| 2 · Slide review (8 frames) | 7,828 | ~800 |
| 3 · Putting it together | ~3,500 | ~500 |
At Opus 4.8's public pricing ($5 per million input tokens, $25 per million output), that works out to about $0.19 for the entire report — versus the better part of an afternoon for a human coach to review the same clip (though a good human coach still catches things the AI can't).
I run these analyses myself, ahead of time — visiting this page never triggers an AI run, so what you see is a stored report. Anywhere you see the marker, the prose next to it was written by the AI coach, word for word.
Rated against Merrill's First Principles of Instruction (5-Star Instructional Design Rating, Merrill 2002) and Mayer's principles of multimedia learning.