Turn any idea, plan, or codebase into a beautiful interactive diagram. An agent skill for Claude Code, Codex, and more.
HTML agent-skillsai-agentsarchitecture-diagramclaude-codeclaude-skillscodexcoding-agentsdeepseek-harnessdeveloper-toolsdiagramsdiagrams-as-codedsh-pluginflowchartllmmermaidopencodesequence-diagramsoftware-architecturesystem-designvisualization
docs(skill): say every ERD fk column needs references in the fast path
tt-a1ia05cca1d5e96
diff --git a/archify/SKILL.md b/archify/SKILL.mdindex 774249f7..7025105a 100644--- a/archify/SKILL.md+++ b/archify/SKILL.md@@ -28,7 +28,7 @@ Use this path for ordinary generation. Read branch references only when their st1. Choose `architecture`, `workflow`, `sequence`, `dataflow`, `lifecycle`, `erd`, `class`, `tree`, `timeline`, or `waterfall` from the question.2. Use the exact schema and example paths in the Type router without listing their directories. Read [Authoring defaults](references/authoring-defaults.md) and the mode's example in a bounded batch separate from project documents and complete schemas so neither is truncated; recover any missing section before writing. For Architecture, use the matching showcase example. For every mode except Architecture, also read the mode and common schemas. Read the relevant schema definition before choosing any new field, enum, or constrained text, especially boundary kinds. Examples teach shape, not facts. Use fresh IDs, wording, and layout. Go directly to the candidate without preliminary help, doctor, starter validation, temporary diagrams, or output-path listing. Query brands only for an explicitly requested mark; read [Brand marks](references/brand-marks.md) for an unknown mark with a user-provided URL.-3. Once the requested scope and, for a real codebase, [source evidence](references/repository-authoring.md) are covered, write the complete candidate directly without planning coordinates in prose. Choose Architecture abstraction and connected placement using Authoring defaults before coordinates: show the main user journey and necessary branches, preserve control roles and behavior-changing conditions, and leave enough room for actual relationship labels. No node, relationship, source, view, card, or boundary count is a target or ceiling. Use automatic routes first; add explicit routing only for necessary branch, return, supplied geometry, or measured repair. Set `meta.quality_profile` to `"showcase"` unless the user requests dense `standard`. For ERD, use a table-first layout: show every physical column in `entities[].attributes`, keep its SQL type and source comment, state every key role a column carries (`key` takes a role or a list, so a junction column is `["pk", "fk"]`), and never move columns into cards as a substitute for the table. Classify tables into functional domains before placement, give each domain a concise `tag`, and place its members in one solid block of grid cells — a single row, a single column, or a full block. A tagged table always needs `row`/`col`, because absolute coordinates have no grid cells and a domain whose tables do not fill one block cannot be drawn as a band: the renderer reports `erd/domain-not-drawn` instead of publishing a diagram that cannot name the domain. A multi-domain schema reads best as one domain per column — each `tag` a vertical run — because the reader scales a wide canvas up to the reader width and shrinks a tall one until the page fits, so the wide arrangement is the one that stays complete and legible inside a laptop viewport. A schema-first ERD over roughly 8 tables or 50 fields is itself a dense map: author its first complete-field pass at `standard` and upgrade only once the layout still reads.+3. Once the requested scope and, for a real codebase, [source evidence](references/repository-authoring.md) are covered, write the complete candidate directly without planning coordinates in prose. Choose Architecture abstraction and connected placement using Authoring defaults before coordinates: show the main user journey and necessary branches, preserve control roles and behavior-changing conditions, and leave enough room for actual relationship labels. No node, relationship, source, view, card, or boundary count is a target or ceiling. Use automatic routes first; add explicit routing only for necessary branch, return, supplied geometry, or measured repair. Set `meta.quality_profile` to `"showcase"` unless the user requests dense `standard`. For ERD, use a table-first layout: show every physical column in `entities[].attributes`, keep its SQL type and source comment, state every key role a column carries (`key` takes a role or a list, so a junction column is `["pk", "fk"]`, and every `fk` column also needs `references` like `"orders.id"`), and never move columns into cards as a substitute for the table. Classify tables into functional domains before placement, give each domain a concise `tag`, and place its members in one solid block of grid cells — a single row, a single column, or a full block. A tagged table always needs `row`/`col`, because absolute coordinates have no grid cells and a domain whose tables do not fill one block cannot be drawn as a band: the renderer reports `erd/domain-not-drawn` instead of publishing a diagram that cannot name the domain. A multi-domain schema reads best as one domain per column — each `tag` a vertical run — because the reader scales a wide canvas up to the reader width and shrinks a tall one until the page fits, so the wide arrangement is the one that stays complete and legible inside a laptop viewport. A schema-first ERD over roughly 8 tables or 50 fields is itself a dense map: author its first complete-field pass at `standard` and upgrade only once the layout still reads.4. Once the complete first candidate is written, run `finalize` directly. Its first gate is showcase validation; successful first drafts need no separate pre-validation. Keep the candidate unchanged while the command runs:```bash