Menü
Create, critique, and validate Mermaid architecture diagrams
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Basierend auf der SOC-Berufsklassifikation
Safely execute a delegated repository task by proving the git environment, protecting dirty worktrees, running required gates, choosing commit/push or patch export, and halting instead of publishing when verification fails.
Use when building or debugging format converters, parsers, or export pipelines (e.g. HTML→PPTX, DOM→IR, doc→doc): three cross-checking techniques — probe-first classification, dual-source extraction, declaration-level validation — because no single representation of the data can be trusted.
Build converter/parser fixture matrices by probing third-party trees before handlers, adding regression fixtures before feature fixtures, declaring expected behavior and primary gates, isolating outputs, and making renderer checks opt-in.
Record decisions reached through interview as a non-authoritative repository document that preserves consensus across sessions and across agents. Use whenever interview outcomes need to outlive the conversation in a form that future readers — both human users and other agents — can act on without re-deriving the reasoning.
Run a structured interview where one decision is taken per turn, automatic inferences from prior consensus are surfaced before manual questions, and depth on each decision is preferred over interview round-trips. Use whenever a user needs to make a sequence of related decisions whose record will outlive the conversation.
Process an external evaluation report into prioritized work groups with clear dependencies, separating procedural fixes from substantive interview questions and keeping each group small enough for a single self-contained delegation. Use whenever an external agent or reviewer has produced a list of findings that the project needs to act on without re-litigating each finding.
| name | mmd-diagram-authoring |
| description | Create, critique, and validate Mermaid architecture diagrams |
| version | 0.1.0 |
| author | html-to-editable-pptx project |
| license | MIT |
| metadata | {"hermes":{"tags":["architecture","diagrams","mermaid","mmd","action-items","repository"],"category":"architecture","related_skills":["consensus-recording","interview-facilitation"]}} |
This skill defines a portable discipline for Mermaid .mmd architecture diagrams. Within html-to-editable-pptx, it also serves as the project's authoring guide: choose action-item keywords first, test adjacent action pairs in natural language, stabilize the macro plot, then decompose accepted actions into lower-level diagrams. The fragile part is diagram judgment and authoring discipline; the only bundled automation is a lightweight .mmd checker.
Good MMD diagrams are not dense drawings: they lock one responsibility level, expose missing consensus, and create a stable path to the next lower level.
When asked to critique or review architecture diagrams — or when macro alignment is unstable — read references/architecture-diagram-evaluation.md first and evaluate against it: consensus-layer checks precede .mmd authoring.
.mmd files.Do not use this for decorative diagrams, one-off unsaved sketches, or diagrams that require a notation other than Mermaid.
.mmd, run the checker, keep en/ko variants aligned. Checker-backed checks live here; these map to the Machine-verified items in Verification, and no design decisions are made in this mode. If a translation or label change requires judgment, switch back to interactive authoring.docs/diagrams/architecture/; the directory must exist before saving.Interactive authoring: pick the diagram layer with the user, draft action-item labels, tune adjacent pairs in natural language, iterate until accepted.
Batch validation: from the repo root, run python skills/architecture/mmd-diagram-authoring/scripts/check_mmd_files.py <target-directory>. If you are already inside the skill directory, drop the leading skill path. Use --log only when you intentionally want a JSONL audit log written to the first directory argument.
| Need | Action |
|---|---|
| Start a new architecture diagram | Pick 5-7 action items, not nouns |
| Check flow naturalness | Read each adjacent pair as a sentence |
| Refine a crowded diagram | Split into macro lifecycle and one subflow at a time |
| Save accepted diagram | Write raw Mermaid to <target-directory>/*.mmd |
| Validate saved files | Run the checker against <target-directory>; add --log only for intentional audit logging |
Clarify the layer. Ask whether the user wants the macro lifecycle or a subflow. Do not mix all layers in one diagram.
Draft action-item keywords. Prefer action labels over nouns. Use patterns such as Register [artifact], Generate [output], or Collect [evidence type]. For project-local examples, consult the label vocabulary reference.
Tune pairs in natural language. For every edge, read the pair as a sentence. If the sentence feels unnatural, rename or reorder the action items before adding detail.
Stabilize the macro plot before micro details. First produce a small lifecycle diagram. Only after the user accepts it, create subflow diagrams for one action at a time.
Name decision owners. Decision nodes should show the actor when it matters, e.g. Reviewer decides: accept or request revision?, not just Accept?.
Preserve lineage. If a flow creates a revised artifact, route it through an explicit action such as Register revised [artifact] or Publish immutable [artifact] revision. Do not jump directly from feedback into a downstream job.
Separate evidence and reports. Keep source analysis evidence, render evidence, validation evidence, and final reports distinct unless the diagram is explicitly about an evidence bundle.
Show concrete outputs. If an action computes a digest, binds a manifest, delivers a PPTX, or publishes a revision, make that output visible in the label or adjacent step.
Save only accepted diagrams. Store raw Mermaid only. Do not include Markdown fences in .mmd files.
Validate .mmd files. Run the bundled checker when available.
This skill stays portable by keeping project decomposition and glossary material in separate references. When you are working inside html-to-editable-pptx, consult:
references/project-diagram-layers.md for this repository's current layer map.references/project-label-vocabulary.md for current action-label examples and the English/Korean glossary.Treat both references as project-local aids, not as generic authoring rules.
Prefer action labels over nouns, and treat any concrete labels as examples rather than as portable standards. If a project maintains a glossary, keep the glossary below the consensus layer and update the skill only when the principle changes.
When producing Korean variants, keep the structure identical to the English file and translate labels consistently. Project-specific translations belong in a glossary or reference file, not in the portable skill body.
Use stable numeric prefixes and language suffixes:
<target-directory>/01_context.en.mmd
<target-directory>/01_context.ko.mmd
Use lowercase snake case after the numeric prefix. Use .en.mmd and .ko.mmd when maintaining parallel language variants.
Bundled project skills live under skills/<category>/<name>/; Hermes exposes
them via skills_list / skill_view / skill_manage. Keep frontmatter
discoverable; details belong to the Hermes documentation, not this skill.
Use the lightweight checker after saving .mmd files:
python skills/architecture/mmd-diagram-authoring/scripts/check_mmd_files.py <target-directory>
The checker validates machine-checkable structure only. Hygiene checks are limited to four things: the first non-whitespace line starts with a supported Mermaid declaration prefix, the file contains no Markdown fences, the file contains no tabs, and the target contains at least one .mmd file. Parity checks compare only same-stem *.en.mmd / *.ko.mmd pairs, and only for node IDs and edge endpoints. Edge labels, display labels, styles, duplicate edge counts, unpaired language-variant files, and files without .en/.ko suffixes are not treated as parity failures. It does not replace human diagram review. Use --log only when you intentionally want a JSONL audit log written to the first directory argument.
v1 and v2 make architecture diagrams look like one execution trace. Use initial, revised, or parameterized terms.Accept? is ambiguous. Write Reviewer decides when user/reviewer judgment is part of the design..mmd files pass scripts/check_mmd_files.py: supported declaration prefix on the first non-whitespace line, no fences, no tabs, non-empty .mmd target sets, and same-stem *.en.mmd / *.ko.mmd parity for node IDs and edge endpoints only