A diagram is not a drawing you maintain. It is a view generated from the model — always current, in whatever notation the reader needs. The rule is directional: read the picture from the source; never transcribe the source into a picture.
Every organisation has architecture diagrams that are quietly wrong. Not because anyone was careless — because a hand-drawn diagram is a copy, and copies drift the moment the thing they describe changes. The fix is not more discipline about redrawing. It is to stop copying.
If structure lives as metadata — entities, relationships, mappings, lineage — then a diagram is just a projection of it. Change the model and every picture changes with it, because there was never a second version to forget. That is the whole argument, and it turns diagramming from maintenance work into generation.
The test: when reality changes, does someone have to open a drawing tool? If yes, you have a picture. If no, you have a view.
There is a reason to care about this beyond tidiness. A structure nobody can see does half its job. Structure makes knowledge exist in a form you can work with; seeing it is what makes it possible to trust, question and act on.
For a data modeller the two were never separate. The core artifact of the trade is a diagram. The conceptual model a business signs off on is a drawing. Structure and its picture are not a thing and a rendering of it — they are the same object, seen at the depth each audience needs.
Each half is weak alone. Structure without visualisation is a model only its author can read. Visualisation without structure is a pretty picture with nothing underneath, drifting the moment the world changes. Together you get the rare and valuable thing: a structure that is real — queryable, validated — and visible, so anyone can read it and push back.
That pushing back is the point. A diagram is where a model meets the people who have to trust it. Make it readable and they will tell you what is wrong — which is how the model gets better.
Three moves, in order. Each is worthless without the one before it.
Capture structure as metadata, not as boxes on a canvas. Entities, relationships, cardinality, ownership — as data. The moment structure is data, everything downstream becomes possible.
Render from that metadata into whatever notation fits: ERD, lineage graph, context diagram, flow. The generator is deterministic, so the same model always yields the same picture.
People still sketch — sketching is thinking. So parse the drawing back into metadata rather than leaving it stranded. The canvas becomes an input, not a dead end.
The order matters. Generating from a model you never built is just a slower way to draw.
Different readers need different pictures of the same thing. An executive needs a context diagram; an engineer needs column-level lineage; an auditor needs to know what feeds what. Hand-drawn, that is four artifacts to maintain. Generated, it is four views of one source.
Text-first, diffable, renders in Markdown and Git. The pragmatic default when the diagram belongs beside the code.
Where people actually sketch. Worth generating into — and worth parsing back out of, so the sketch is not a dead end.
Layout engines that do the hard part. You describe what connects to what; they decide where it goes.
For the large graph nobody can hand-arrange — hundreds of entities, laid out automatically and still readable.
Notation is a rendering choice, not an architecture decision. If switching from Mermaid to PlantUML means redrawing everything, the diagram was never generated in the first place.
“Diagram” is not one thing. Each type answers a different question, and picking the wrong one is why so many diagrams get ignored. All of these are generated from the same metadata — what changes is which slice of the model is projected, and how.
Answers: what things exist, and how are they related? The oldest and most misused type. An ERD of 200 tables is not a diagram, it is wallpaper — the discipline is choosing the subject area first and letting the generator draw only that. Conceptual, logical and physical are three different ERDs of the same reality, and mixing them is the commonest way to lose an audience.
Concepts: Living Diagrams · Business-Friendly Diagramming | Articles: Generating yEd diagrams from metadata
Answers: what feeds this, and what breaks if I change it? Two flavours worth separating: horizontal lineage follows data across systems; vertical lineage follows a single column down through its transformations. Auditors want the first, engineers want the second, and a diagram that tries to be both is usually neither.
Concepts: Business-Friendly Data Lineage · Horizontal vs Vertical Lineage | Articles: Dynamic Metadata Lineage
Answers: in what order does work happen, and what depends on what? A DAG is lineage with execution semantics attached — the arrows carry runs after, not just flows into. Generated from orchestration metadata rather than drawn, it becomes the one picture that is true at 3am.
The acyclic part is the constraint that makes it useful: a cycle in a DAG is a bug, and drawing it exposes the bug.
Answers: how do concepts relate across domains? Where an ERD is bounded by a schema, a knowledge graph crosses them — entities, decisions, documents, people, all in one connected structure. Its shape is a triple: source · relationship · target. That is also the shape of a link table, which is why a vault of notes and an enterprise ontology converge on the same architecture.
Concepts: Conceptual Constellations | Research: The visual tool landscape
Answers: what is this system, and where does it sit? Fewest boxes, most deliberation. A context diagram is almost entirely about what you leave out — which is exactly the human judgement a generator cannot make for you.
Concepts: Generated Palettes · Drawing It Together
One model, several types. The same metadata produces an ERD for the modeller, a lineage graph for the auditor, a DAG for the engineer and a context diagram for the executive. Hand-drawn that is four artifacts drifting apart. Generated, they cannot contradict each other.
The named ideas behind the visual layer — Living Diagrams, Drawing It Together, Visual Thinking, Clickable Diagrams — and the anti-pattern they all argue against: Model as Drawing.
Each is a small, self-contained idea you can point at in a meeting. Together they make the case that a diagram is an output, not an artifact.
This argument was not invented for this site. It has been worked out in public across two dozen pieces since September 2025 — and they read in an order: choose an approach → generate → close the loop → follow the meaning → publish and keep it true.
Before generating anything, decide what you are generating into — and on what grounds.
The mechanics, per notation. Each is a small generator, not a platform.
.drawio files. coming soonPeople sketch. The question is whether the sketch is a dead end.
Once structure is data, the picture stops being an ERD and starts answering questions.
A generated diagram nobody sees is still a diagram nobody sees.
Articles publish on Medium and Structure Beats Magic. Unlinked pieces are written and queued — they appear here as they go live.
Not everything with a name is a concept. Most of the vocabulary here is industry language we use consistently — worth defining once so a reader is not guessing, but not ours.
Field vocabulary anyone in data would recognise: ERD, DAG, impact analysis, round-trip, blast radius. We clarify it and link it. We did not invent it.
A named idea this work coined and owns: Living Diagrams, Generated Palettes, Model as Drawing. You would not find it in someone else’s glossary.
The layer is deliberately volatile. A term that keeps recurring, earns its own article, or grows a distinctive angle gets promoted to a concept. And the reverse: something we treated as a concept, which turns out to be standard vocabulary, moves back down. Being honest about which is which is the point — a glossary padded with borrowed terms dressed as original ideas helps nobody.
Questions investigated properly and written down — sources checked against the primary record, findings dated, open gaps stated rather than smoothed over. Each is a markdown file with a metadata header; this section is generated from those files.
Why publish research at all? Because studying how other tools solve the visual problem is not idle curiosity — it is requirements-gathering with evidence. Write the findings as structured metadata and a specified feature is already halfway built.
Claude skills are folders with a SKILL.md: frontmatter plus instructions, loaded on demand. They are the obvious way to make a diagram generator available inside a conversation — so we went looking for the ones that already exist.
The finding was thinner than expected. Anthropic’s official repository holds 17 skills and not one generates a diagram. A GitHub search for diagram skills returns two repositories, with 2 and 0 stars. Two of the three skills most commonly cited in write-ups return 404.
There is no ERD skill. For anyone modelling data, that is an unoccupied space — and the reason is instructive: the visual skills that do exist draw from a prompt. A useful one would generate from the model: deterministic, multi-notation, carrying provenance, and able to round-trip. Identical in a demo; different in month three.
Skills we intend to publish: an ERD skill (metadata in; Mermaid, draw.io and GraphML out) and a brief skill (markdown + frontmatter as an installable convention). Both are packaging work — the generators already exist.
The people and tools worth following on the visual side. Curated, not collected — each is here for a stated reason, including the ones that argue against this site’s position.
Zsolt Vicziá́n — maker of the Obsidian Excalidraw plugin. The rare source that covers both the format (open, generatable JSON) and the method (sketching to think). Most relevant of the four.
Visual facilitation and visual minutes: what happens in the room when a group draws together. The human half of modelling, practised as a craft. Author of Visual Thinking, Visual Doing, Visual Working and My Icon Library. Trained there — see News.
Doug Neill on sketchnoting and visual journals. Furthest from metadata, closest to the reason any of this works: drawing while listening is how understanding forms. Quiet since March 2026.
Included as the honest counter-argument: “I don’t want to be in control of my ideas.” Structuring too early kills connections you had not yet seen. Worth answering, not dismissing.
Also tracked: Heptabase (card-as-unit, board-as-view) and Vault-LD (notes as linked data).
Willemien Brand’s four titles (BIS Publishers) are the most accessible route into visual working:
The foundation — why drawing improves thinking, and why everyone can do it.
Applying it: from idea to something other people can see and respond to.
Business drawing as a daily routine, step by step.
Building a personal icon vocabulary. Conceptually the closest to this site: a library assembled once and reused endlessly — the same idea as a generated palette, produced by hand.
Zsolt Vicziá́n’s workshop, starting March 2023: Excalidraw in Obsidian, and the book-on-a-page method — compressing an entire book into a single visual summary. Forces the question every model has to answer: what actually matters here? March 2023.
Den Haag, with a follow-up Q&A. Willemien Brand’s facilitation craft first-hand — how a group draws its way from conversation to decision. The counterweight to everything generated: some pictures have to be made by people, in a room, together. October 2021.
The visual layer of Breakthrough Modeling gets its own home — 25 articles, a concept library, and research on the tool landscape. 31 July 2026.
Two trainings, two halves of the same craft: one on compressing meaning into a single page, one on drawing it out of a room. Neither is generated — and that is the point.
This is the visual layer of Breakthrough Modeling. Structured Metadata is the model — what the structure is. This site is the rendering — how you show it, to whom, in which notation.