# Format Layout Diagram > Draw the layout of a memory or wire format that already exists in code — a byte anatomy of one encoded instance, the per-variant conventions of a shared encoding, or the linkage between blocks. Use when asked to "diagram the format", "show the byte layout", "draw the header", "picture the on-disk structure", "explain the encoding", or when a format reference, design record, or pull request needs a figure of a structure a reader cannot hold in their head. Beta — the conventions come from one format in one repository, so record friction and confirmation through `engineering-journal`. Symptoms that this skill applies: a byte grid drawn by hand from a specification, a figure whose offsets no longer match the encoder, a byte grid routed through a graph layout engine, a figure that cannot say whether it shows what ships or what is planned, or a figure that looks right in a browser and breaks in the rasterizer that publishes it. - Skill: `iopsystems/format-layout-diagram` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add iopsystems/format-layout-diagram` - Raw SKILL.md: https://api.skillmd.com/api/skills/iopsystems/format-layout-diagram/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: iopsystems (https://skillmd.com/u/iopsystems) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/iopsystems/format-layout-diagram --- # Format layout diagrams A layout figure answers a question prose answers badly: **what occupies which bytes, and what addresses what.** Prose carries a sequence; a layout is a shape with addresses, and re-deriving it from sentences is the work the figure exists to remove. This skill is the third in a family and inherits its principles: derive-never-draw, fail-loud, verify-the-rendering. `architecture-diagram` owns what exists and what contains what; `dataflow-diagram` owns what moves where. When the question is "which byte is that field in", it is this one. ## What this is, and what it is not yet Treat this skill as **beta**. Its conventions come from one binary collection format in one repository, whose figures were built without a skill and whose author proposed this one afterward. Three rules below are that author's, stated in their own words; the rest is inference from the artifacts and carries less weight. When one fights your format, the override and its reason are the most valuable thing the effort produces: record it through `engineering-journal`. ## Three rules **Position is the byte offset. No layout engine.** In a byte grid the x-axis *is* the address, so nothing may be free to move a cell. Emit the drawing directly, with integer geometry and fixed ordering. Routing a byte grid through a graph layout engine was tried and abandoned: clusters and edges moved cells, which broke the one claim the figure exists to make. **Spans come from golden fixtures decoded by the shipping codec.** Not from the specification, and not from a decoder written for the figure. A second decoder is a second source of truth — one was written, then deleted for that reason. The generator walks a frozen test fixture through the production decoder and aborts when any drawn span disagrees with the bytes. **Freshness is a deterministic regenerate-and-diff on the artifact readers see.** Not a checksum of the source, and not a promise to remember. Because emission is deterministic, regenerating and diffing covers exactly the file a reader opens. ## The derivation contract A derived figure is produced by a program that fails loudly rather than drawing something plausible: - **Assert every span against the bytes.** A length that does not match, a count that does not match the walk, a stored offset that does not land on the element it names — each aborts, naming what disagreed. - **Re-encode and compare.** Decoding proves the reading; re-encoding each decoded value and comparing bytes proves the figure's labels describe a value the codec would produce. It catches a tier table that has drifted from the encoder, which decoding alone does not. - **Check the frame, not only the content.** Cells past the drawing bounds are an abort, not a cosmetic problem. A figure that silently clips is worse than one that refuses to render. - **Freeze the instance.** Draw a fixture the tests already pin, and name it. - **State the regeneration command** in the artifact's own header and in the provenance table, so the next person changes the generator rather than the output. Choose the instance for coverage, not size: the smallest encoding that exercises the header, at least two entries, at least two width tiers, and every cross-reference the format defines. A one-entry example teaches the header and nothing else. ## Source format follows what the figure claims A byte grid and a relationship schematic are different kinds of claim, so they keep different sources. Record the choice in a provenance table beside the figures — one row per figure, naming its source, its regeneration command, and what it asserts. - **Byte anatomy — generated, emitted directly, no render step.** The committed artifact is exactly what the generator produced. A render step between generator and artifact reintroduces a layout engine's freedom over a grid whose positions are addresses. - **Relationship schematic — graph source, rendered.** Chaining, linkage, shipped-against-reserved. Here a layout engine is correct: the claim is what connects to what, not where anything sits. The graph source is the diffable artifact; its render depends on the renderer's version and is refreshed on edit. - **Hand-authored snapshot — direct drawing, no generator.** Variant conventions and other design claims that no code produces. Date it and name the review it came from, because nothing will detect its drift. A generated figure may not draw what does not exist. A schematic may, if the unbuilt half is visually distinct — dashed for reserved, solid for shipped — and the caption says which is which. ## Verify the rendering, not the markup A figure that is valid and unreadable has failed. Rasterize the committed artifact before committing it, with the same tooling the publishing host uses. Two defects found this way in one figure, neither visible in a browser: a double hyphen inside an XML comment, which is illegal and rejected by strict parsers; and a root style setting maximum width with automatic height, which collapsed the canvas to zero height in the rasterizer while looking correct locally. Rasterizing proves the figure drew, not that it is balanced. A bracket label nearer a neighboring span than its own, a two-line label centered by its first baseline, an offset arrow attached wherever the code reached first — each is inside the frame and each is what a reader sees first. `architecture-diagram` carries those placement rules, which a byte grid inherits the moment it emits its own geometry. When the rasterizer a check depends on is unavailable, say so and name what stood in for it — markup validity, a deterministic regenerate compared by hash, bounds and collision checks, a reviewer reading the committed artifact. A check skipped without that record reads afterwards as a check passed. Also: emit fixed ink over an opaque ground rather than colors inherited from the page, because a host that embeds the figure as an image strips that context. Give the figure a text alternative twice — an accessible label inside it, and prose beside it stating the same claim. ## A default visual language Adopt wholesale; override with a reason. - **One cell per byte**, fixed width, in offset order, value in hex. Bytes are the unit even when fields are wider, because the reader's question is "which byte". - **Fill by role, not by field** — header, tag, payload, trailer — reused across every entry. Four colors learned once beats one per field. - **An offset ruler under the cells**, at reduced opacity: a reference the eye returns to, not content. - **Brackets above for spans**, staggered when labels would collide, each carrying the field name and its decoded value. The value is what lets a reader check the figure against the bytes. - **An arrow for every stored offset**, from the field to the element it addresses. This is the part readers remember, because it turns a number in a header into a place in the picture. - **A caption that states the claim**, not the contents. "The tail offset points at the last entry's first byte, so tail access is constant time" earns its space; "diagram of a block" does not. ## Boundaries `architecture-diagram` owns structure; `dataflow-diagram` owns movement. A reader asking which byte holds a length is asking neither. `technical-prose` owns wording and scale in captions and surrounding prose, including its default of American spelling. Field names, type names, and byte values are untouchable — they are names, and a figure that renames them to read better has stopped describing the format. `engineering-journal` owns the durable record of why a format is shaped as it is. The figure carries the shape; the entry carries the argument. ## Known dead ends - **A graph layout engine for a byte grid.** Clusters and edges move cells, which breaks position-is-offset. Emit directly instead. - **A second decoder written to drive the figure.** It is a second source of truth, and the two drift. Derive through the shipping codec. - **Drawing the grid by hand from the specification.** Correct on the day it is drawn, wrong at the first encoder change, with nothing to detect it. - **Freshness by convention.** "Regenerate when the format changes" is a promise, not a check. Until the regenerate-and-diff runs somewhere, assume the figure is stale. - **One figure for the shipped and the planned shape, undifferentiated.** Readers act on all of it as though it exists. ## Red flags - A byte grid whose cell positions came from a layout engine. - A byte grid with no generator, or a generator with no abort path. - A generator that decodes but never re-encodes to compare. - A figure with no provenance row: no source, no regeneration command, no statement of what it asserts. - A committed artifact that is a render of a generated intermediate rather than the generator's own output. - Offsets that no longer match the encoder, with nothing that would have caught it. - A fixture that is not the one the tests pin. - Cells past the drawing bounds, silently clipped. - A bracket or arrow label nearer a neighboring span than the one it names. - A rasterization check skipped for unavailable tooling, with no record of what stood in for it. - An artifact committed without being rasterized the way the host will. - A figure that inherits its colors from the reader's theme. - A schematic that omits which half is reserved. - A caption that names the contents instead of stating the claim. - A field name rewritten in a label so the figure reads better.