Case study
Physics of Connection: how to share a thought
Improv, taught as the mechanics of communication: typed concepts, lessons as paths through them, and a first video that held viewers for 74% of its length.
- average view duration on the first lesson video (3:12 of 4:18)
- 74%average view duration on the first lesson video (3:12 of 4:18)
Figures measured 2026-10-13.
The problem
Improv carries some of the most practical communication training there is (listening, reading a room, building on what someone else offers), but it is taught in rooms, by feel, and written down as scattered exercise lists and memoir. Someone with a specific problem, like overthinking in conversation, has no way in.
What was built
Modelled the subject as a knowledge graph rather than a blog. Concepts are typed (law, technique, exercise, principle, antipattern and more) and the type decides where a concept is published. Concepts link through explicit relations (requires, enables, contrasts, extends, illustrates). Lessons are threads through concepts; learning paths sequence lessons.
Navigation starts from the reader’s problem rather than the subject’s table of contents: pick a symptom such as overthinking or awkwardness and the graph returns the concepts and practice that address it.
Prose is auto-linked to the graph at render time, with deliberate exceptions for ambiguous one-word concepts, and tests enforce the content rules, such as every search alias actually appearing in the text it names.
Interactive practice tools (an exercise picker, a prompt generator), and lesson videos rendered in code with Remotion rather than edited by hand.
Search targeting with a hard rule: every keyword figure comes from measured data, and a field with no measurement is left empty rather than estimated.
What changed
A live product a reader can enter from a problem and leave with a practice routine.
The first lesson video held viewers for an average of 3:12 of 4:18. That is a strong attention signal for a channel with no audience yet; reach is the open problem, and the write-up says so.
Inside your team
Structuring knowledge as typed, linked units is exactly what an AI system needs to retrieve the right context, and what an enablement program needs to teach a team. The same modelling turns a company’s scattered playbooks into something both people and agents can use.