How this course works
Six sessions, 45 minutes each, and one dataset the whole way through: Daybreak's subscription numbers. Same numbers every session, so the only thing that changes between one chart and the next is a decision you made.
★ What you walk out with tonight. A written question for a chart you already own, its job named from a list of five, and at least one chart you are going to replace with a sentence.
LiveWhy almost everything in this course is checkable2 min▶
Data visualization gets discussed as taste, and most of it is not. Whether a picture overstates its own data is arithmetic. Whether two series colours can be told apart by a colour-blind reader is arithmetic. Whether your axis labels are readable is arithmetic.
Session 3's bench computes all three from the chart it is currently drawing. That is unusual for a course, and it exists to make one point: the things people argue about in chart reviews are mostly not the things that matter, and the things that matter can be settled with a number.
What is left after the arithmetic is genuine judgement - which question, which form, what to leave out. That is what sessions 1 and 2 are for, and no tool will do it for you.
Five jobs, and the form follows from which one it is
6 min liveNot five chart types - five things a reader might be trying to do. Name the job and the shortlist of forms is two long instead of twenty.
LiveWhen it should not be a chart at all4 min▶
Three situations, and all three are common:
- One number. A chart of one value is a number wearing a costume. Set it large, put the comparison next to it in words, and you are done.
- Two numbers. "Up from 310 to 406" is a sentence anybody can read at a glance. A two-bar chart makes them look at a picture in order to get back to the two numbers.
- Exact values matter more than the shape. If the reader will be quoting the figures, they want a table. A table is a legitimate visualization and treating it as a failure is why so many reports contain charts nobody can read numbers off.
The test: can you say the finding in one sentence, and does the sentence lose anything? If it does not lose anything, ship the sentence. You will save the reader the decoding and yourself the afternoon.
A two-bar chart: "Cancellations, last month and this month." Axis from 0 to 450, two bars, a legend, a title, gridlines.
The reader looks at it, works out which bar is which, follows the taller one down to the axis, estimates 400-ish, does the same for the other, and arrives at "up by about a third". Six seconds and an estimate, from a chart that contained two exact numbers all along.
Cancellations: 406 this month, up from 310. That is 31 percent, and it is the third consecutive rise.
Read at a glance, exact, and it carries a piece of context the chart did not. It took no design decisions, no colour, no axis. If somebody then asks how the rise is shaped over the year, that is a chart - and now it is a chart with a question behind it.
A chart answering three questions answers none of them
6 min liveThe pressure to combine is enormous - one chart is less work to build, less space on the page, one thing to explain. It is also the most reliable way to make a chart nobody uses.
LiveWriting the question down, and what it decides4 min▶
One sentence, above the chart, before you build it. It is the same discipline as an objective in learn-course-design-with-phoebe or a problem statement in learn-product-design-with-phoebe, and it decides four things at once:
| The question | What it fixes |
|---|---|
| "Which roast earns most in a month?" | Job: magnitude. Form: sorted bar. One series, one colour. Axis from zero |
| "Is the subscriber base growing?" | Job: change. Form: line. Axis need not start at zero, and session 3 explains why that is fine here and not there |
| "Which cohorts are churning fastest?" | Job: comparison across two dimensions. Form: heatmap or small multiples |
Notice that the third column is not a matter of preference in any of those rows. The question determined it.
Once you have written the question, the chart's title is its answer: not "Revenue by roast" but "Ethiopia Guji earns most, by 21 percent." A title that states a category is a filing label. A title that states the finding is the chart doing its job before anybody looks at it.
That seeded title is a filing label: it says what is in the chart, not what the chart found. It also names two dimensions and a three-year window, which is the tell for a chart carrying more than one question. The finding version - "Ethiopia Guji earns most, by 21 percent" - is shorter and does the reader's work for them.
Name the question, name the job, and cut one chart
★ 22 min · everyone buildsNo drawing tonight. Five short decisions, made in writing, about charts that already exist.
The question. One sentence, in the reader's words, that the chart exists to answer. If you cannot write it, that is the finding - you have a dataset and a slide slot, not a chart.
The job. Magnitude, change, identity, polarity, or one number. Pick exactly one. Two jobs means two charts.
Not a chart? Say the finding out loud as a sentence. If nothing is lost, ship the sentence and move on with your afternoon.
One or three. Count the questions your draft is carrying. If it is more than one, split it - three simple charts cost the reader less than one complicated one.
The title. Rewrite it from a label into the finding. "Ethiopia Guji earns most, by 21 percent."
Open a real chart you made in the last month. A slide, a dashboard tile, a screenshot in a document. Real ones only - the friction is the exercise.
Write its question in one sentence. Then check it against the chart. A surprising number of real charts answer a slightly different question from the one their maker would state.
Name the job from the five. If you find yourself wanting two, you have found a chart that should be two charts.
Say the finding out loud. "Cancellations are up 31 percent, the third rise in a row." Did the chart add anything to that sentence? Be honest - roughly one chart in four does not.
Rewrite the title as the finding. This costs thirty seconds and it is the single highest-value change most people can make to their existing charts, because it works even on the badly built ones.
Now do it for the Daybreak set. Signups by month, revenue by roast, retention by cohort, cancellations by reason. Four questions, four jobs - and decide which of the four should be a sentence.
Write your not-this line. One thing this chart deliberately does not show. It is what stops the request to add a second axis in week three.
Notice the last row is marked as a trap and is not one. Splitting the overloaded chart adds two to the count, and it is the right call every time. This is the honest limit of counting things: fewer charts is usually better and it is not the goal. The goal is one question per chart, which sometimes costs you two more slides.
Before session 2
20 min- Run the five decisions over one real deck. Count how many charts survive as charts. A quarter becoming sentences is normal.
- Rewrite three titles from labels into findings. Send one of them out and watch whether anybody asks you what the chart shows.
- Find your overloaded chart. Every team has one. Count the questions in it.
- Find a table pretending to be a chart - somewhere the reader clearly needs exact values and is being given a shape instead.
- Bring one chart to session 2 whose form you are not sure about. Session 2 is the shortlist.
Three questions
Clear all three and session 1 gets stamped in your passport.
1 · When should a finding not be a chart?
A chart of one value is a number wearing a costume, and a two-bar chart asks the reader to decode a picture in order to recover two numbers they could have read. A table is a legitimate visualization, and treating it as a failure is why reports are full of charts nobody can read figures off.
2 · Your draft shows signups and cancellations as lines with revenue as bars behind them, split by roast. What is the problem?
Combining feels efficient because it is one chart to build and one slot on the page. It is not more informative - it contains the same facts in a form that costs more to read. Three simple charts fit in less attention than one complicated one.
3 · Which is the better chart title?
A title that names categories is a filing label - it tells the reader what is in the chart and leaves the work to them. A title that states the finding does the chart's job before anybody looks at it, and it costs thirty seconds. It also works on charts that are otherwise badly built, which makes it the highest-value retrofit available.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map. Daybreak's numbers are a teaching dataset.
Session 1 cheat sheet · pin this
You have a question and a job. Session 2 turns that into a shape: bar, line, dot or heatmap chosen from the job rather than from habit, sorting as the cheapest improvement available, small multiples instead of eight series - and the one rule in this course with no exceptions, which is that there is never a good reason for two y-axes.