The research

1. 1Larkin and Simon, 1987Text makes the reader search for the facts and infer how they connect. A diagram works by locality: what belongs together sits together, so the reader looks instead of searching.
   In ttmi

   • →The layout keeps connected parts next to each other, one column per step along the arrows.
   • →A note sits beside the part it explains: a note that only points at one shape is placed right next to it.
2. 2Glenberg, Wilkinson and Epstein, 1982Readers miss contradictions in text they have just read, and still rate their understanding as high.
   In ttmi

   • →The explanation is a picture you can check at a glance. A wrong part or a missing arrow is plain to see.
   • →You can comment on the part that looks wrong and hand the sketch back to your agent, which answers on the board.
3. 3Bransford and Johnson, 1972The same passage was hard to understand and remember without the picture it described, and easy with it, but only when the picture came first.
   In ttmi

   • →A walkthrough opens on the whole drawing for a moment before the first stop, so every stop after it lands on a picture you already have.
   • →The drawing stays in full for the whole walkthrough. The camera does the focusing, so you stay inside the final picture rather than watching it being built.

1. 4Heiser and Tversky, 2006With arrows a drawing reads as cause and function; without them, as structure. Lines are relations, and arrows are asymmetric.
   In ttmi

   • →Every arrow between parts carries a label: a verb or the thing that moves, four words at most.
   • →A note's connector is drawn as a plain line, because "this note describes that part" is a relation, not a cause. It is worked out from where the line starts, so dragging its end off the note turns it back into an arrow.
2. 5Moody, The Physics of Notations, 2009One symbol, one meaning. Few symbols. Text plus graphics. Hierarchy to manage complexity.
   In ttmi

   • →A sketch is three representations and no fourth: the diagram for structure, code for mechanism, steps for time.
   • →Fill is visual weight: solid for the one or two parts the sketch is about, tint for the parts that act, none for context.
   • →Colour means one thing per sketch, and a legend says what. Red always means failure or danger. A tag says what kind of thing a part is, so colour is never spent on kinds.
   • →Groups are the hierarchy.
3. 6Sweller, cognitive loadA worked example beats an abstract description, and detail placed away from the thing it is about splits attention.
   In ttmi

   • →A note shows a type plus one example value, not the type alone, and a mechanism as a few lines of pseudo code.
   • →The note sits next to the part it is about, joined to it by a line.
   • →Agents check a budget before saving: at most 12 shapes, one numbered path of at most 7 arrows, one note per stop of at most ten lines, at most four groups. A bigger picture becomes several sketches, the overview linking to the detail.

1. 7Shneiderman, 1996Overview first, zoom and filter, then details on demand.
   In ttmi

   • →Notes are folded to their heading in the overview, and the layout uses the folded size, so the structure fits one screen.
   • →A note opens when you click it, or when the narration names it.
2. 8Palmer, Common region, 1992A shared region groups things harder than closeness or similarity. Whatever lands inside a region is read as a member, and two regions that cross read as shared membership.
   In ttmi

   • →A group's region holds only its members. The layout keeps every other part out of it, rather than drawing a region that makes a false statement.
   • →A group is drawn around its member boxes only; a note hangs beside its box outside the region.
3. 9Sander, 1996; Rodgers, Zhang and Purchase, 2012Layered layouts with clusters keep each cluster in one run of slots in every layer, so clusters never cross. Regions read best as simple, separate shapes, never sharing a border.
   In ttmi

   • →A group is a lane that keeps its slot in every column it spans, so its region is one rectangle that touches no other.
   • →The guide asks for two to four groups, and one kind of boundary per sketch: systems, or teams, or machines.
4. 10Eades, Lin and Smyth, 1993Before laying out a graph in layers, set aside the arrows that close a cycle, chosen by an ordering that starts where the flow starts.
   In ttmi

   • →A feedback loop reads in the order of its story, with one arrow going back, instead of with its start pushed to the end.

1. 11Dyson and Kipping, 1998; Dyson and Haselgrove, 2001Lines of about 55 characters are the best trade-off between reading speed and understanding, and what readers prefer. 25 is slow; 100 is fast but disliked.
   In ttmi

   • →A note is 420 pixels wide, about 60 characters of text.
   • →Code in a note is written in lines under 60 characters, and never wrapped.
   • →The narration caption is at most 600 pixels wide.
2. 12Trollip, 1986, and reading by fixationsJustified text reads slower than text aligned left with a ragged edge. A label is read as a unit when it fits in one or two fixations; a label that wraps becomes text to read.
   In ttmi

   • →The caption is aligned left.
   • →A box label is one line, or two short ones. An arrow label is one line of four words.

1. 13Tversky, Morrison and Bétrancourt, 2002Discrete steps beat animation. What is shown has to be perceivable in the time it is shown.
   In ttmi

   • →A walkthrough is a series of stops, not an animation. At each one the camera glides to the parts being talked about.
2. 14Mayer, segmentingExplanations cut into segments the learner moves through at their own pace teach better, most of all for complex material.
   In ttmi

   • →You press Next. Nothing is timed between stops, so that time is yours.
3. 15Mayer, signallingPointing out the part that matters right now helps the learner pick it out of the whole.
   In ttmi

   • →The arrows of the path carry numbered badges, 1, 2, 3, so the order shows even when you are not playing.
   • →At each stop only that arrow's number turns to the accent colour, and the shape the voice names lights up as it is said.
   • →Dimming everything else was tried and felt wrong in use; the camera does the focusing instead.
4. 16Mayer, spatial contiguityWords belong next to the picture they describe, not across the screen from it.
   In ttmi

   • →While playing, the narration is a caption at the bottom, and the camera centres the parts in the space just above it.

1. 17Mayer, modalitySpoken words beside a picture teach better than written ones, because the eyes stay on the picture.
   In ttmi

   • →Every stop is spoken. The model runs on your device, so the words are never sent anywhere.
2. 18Mayer, redundancyA picture with narration teaches better than a picture with narration and the same words on screen: reading pulls the eyes off the picture.
   In ttmi

   • →With the voice on, the caption is quiet: the controls and a button to show the words if you want them. With the voice off, the words are the narration.
3. 19Mayer, personalizationA conversational style matters more than any timing, and modern synthetic voices teach about as well as human ones.
   In ttmi

   • →The guide asks agents to write for the ear: what they would say to a colleague at a whiteboard, one idea per sentence, talking to you.
4. 20Elmers and others, 2021Pauses at the boundaries of speech help recall at half a second, but not at a fifth of one.
   In ttmi

   • →The voice model's own pauses measured about a quarter of a second, so ttmi holds them longer: about half a second after a sentence.
   • →When a stop names a shape for the first time, the voice holds after it, so your eye can find what just lit up.
5. 21Wingfield, Peelle and Grossman, 2003A faster speech rate and more complex sentences make each other harder to follow.
   In ttmi

   • →Each stop is spoken a little slower the more it brings in and the denser its words. You can set your own speed on top.
6. 22Fraundorf and Watson, 2011Fillers like "um" help listeners remember human speech, but showed no clear effect in synthetic speech.
   In ttmi

   • →Agents write no fillers, pauses or emphasis. They write the words, and ttmi times them from what each stop brings in, the way it lays out the shapes.

• One stop, one breathThe voice model shapes a sentence's melody from everything it is given at once. Cut into pieces, a stop sounded artificial, so each stop is spoken in one go, after the last sentence of the stop before, so it never starts cold.
• Said the way your team says itSmall voice models mispronounce jargon. ttmi gives the voice the sounds it was trained on, and lets you say how your product names, acronyms and terms are pronounced.
• Why this voiceReviews of the small voice models that run in a browser, in October 2026, found newer ones to be different trade-offs rather than replacements, and none that documents word timing. Kokoro does, which is what lets a shape light up as its name is said.

References

• Bransford, J. D. and Johnson, M. K. (1972). Contextual prerequisites for understanding.
• Dyson, M. C. and Haselgrove, M. (2001). The influence of reading speed and line length on the effectiveness of reading from screen.
• Dyson, M. C. and Kipping, G. J. (1998). The effects of line length and method of movement on patterns of reading from screen.
• Eades, P., Lin, X. and Smyth, W. F. (1993). A fast and effective heuristic for the feedback arc set problem.
• Elmers, M. and others (2021), on the length of pauses at speech boundaries and recall.
• Fraundorf, S. H. and Watson, D. G. (2011). The disfluent discourse: effects of filled pauses on recall.
• Glenberg, A. M., Wilkinson, A. C. and Epstein, W. (1982). The illusion of knowing: failure in the self-assessment of comprehension.
• Heiser, J. and Tversky, B. (2006). Arrows in comprehending and producing mechanical diagrams.
• Larkin, J. H. and Simon, H. A. (1987). Why a diagram is (sometimes) worth ten thousand words.
• Mayer, R. E. Multimedia learning: the segmenting, signalling, spatial contiguity, modality, redundancy and personalization principles.
• Moody, D. L. (2009). The "physics" of notations.
• Palmer, S. E. (1992). Common region: a new principle of perceptual grouping.
• Rodgers, P., Zhang, L. and Purchase, H. (2012). Wellformedness properties in Euler diagrams: which should be used?
• Sander, G. (1996). Layout of compound directed graphs.
• Shneiderman, B. (1996). The eyes have it: a task by data type taxonomy for information visualizations.
• Sweller, J. Cognitive load theory: the worked-example and split-attention effects.
• Trollip, S. R. and Sales, G. (1986). Readability of computer-generated fill-justified text.
• Tversky, B., Morrison, J. B. and Bétrancourt, M. (2002). Animation: can it facilitate?
• Wingfield, A., Peelle, J. E. and Grossman, M. (2003). Speech rate and syntactic complexity as multiplicative factors in speech comprehension by young and older adults.