Eight stable diagram types, each keyed to a data shape rather than a topic. A new diagram joins the family whose shape it carries; a new shape earns a new family.
components + connections + membership
What is this system made of, and who talks to whom?
A system drawn as data: nodes on a grid, edges between ids, regions declared by membership. One packet walks the request path, but the walk is DECORATION — the structure is the content.
Edge stackThis siteAWS serverlessKubernetesAI agentAgent harnessAgent anatomyVideo studioProject matrixConcept filmMusk loop
one workload × two designs
Where exactly do two ways of doing the same job part company?
Two pipelines side by side, the SAME input flowing through both. The argument is the divergence point — if you cannot name the shared input and the step where the designs split, it is not a versus.
Codex vs driftRebrand vs tokenInstall vs describeFile vs repoStuff it vs edit itRAG vs CAGSelf-review vs external check
n jobs × m vocabularies
What is this called in each of these systems — and where does the translation break?
A translation table whose rows are jobs no column owns. The exact cells are the on-ramp; the content is the mismatch — a smaller claim, another model, or a blank where a platform has decided the job is not its job. Against Versus: no pipeline and no winner, and more than two columns; against Catalog: remove a column and every row reads differently.
Cloud servicesCLI commands
n named concepts × one shared clock
What kinds exist in this field, and what does each look like?
A field guide: N independent panels, each replaying its own micro-story, paced by one clock so the poster reads as a single organism. Removing any panel must not break the others — if the panels depend on each other, it is a run, not a catalog.
Prompt leversHandback momentsPlausible failuresOwned sourceRepo powersAgent patternsLoop exits10 evals15 internalsDeployment
ordered stages × causal transitions
How did this field get here — and why was each step forced?
A chain whose nodes replace each other in time and whose EDGES carry the argument: every transition names the ceiling that forced the next stage. Against Architecture: lineage nodes never coexist — the arrows are causation, not traffic; against Catalog: remove a stage and the chain breaks.
Agent engineeringAI fluency
n independent axes × one position on each
Where is this subject lopsided — and what is the shortfall costing?
Several dimensions that do NOT order each other, and a subject whose content is the SHAPE they make together. Against Lineage: if being high on one axis implies the others, it is a ladder; against Catalog: catalog panels never compose into one reading. The claim usually lives in the shortfall, not the maximum.
AI fluency
ordered stages × owners, one object's state
Who does what, in what order — and where does the work change hands?
A sparse grid: columns are the phases, rows are the parties that own them, and one object is threaded through with every step a transition of its state. Both axes are data — move a step to another lane and the diagram says something false. Against Lineage: drop the owners and ordered stages with causal edges is already a lineage; against Runs: a run's node positions are layout, and a run has no owners to hand off between. The reading is the CROSSING, and a process that cannot fail is a chain, not a member.
Order fulfilmentIncident responseExpense claimLoan originationContent publishingRelease change
a structure + one walk over it
How does a single execution or traversal flow through this structure?
A graph of named nodes where the WALK is the content: counts accumulate hop by hop, stages change, meters accrue. The test against Architecture is exactly this — if pausing the walk empties the diagram of its point, it is a run.
LoopGraphDynamic workflowsAgent runExecution timelineDecisionsHandoffEgress firewallCRM workflowsCRM data modelCRM access model