AnyCap Architecture Diagrams
Turn a system design into a clean architecture figure with AnyCap. Distilled from repeated real use; the loop below is the one that converges.
0. Sanitize the fact graph FIRST (public-safety pass)
An architecture diagram is a浓缩的系统地图 — it leaks exactly what scans can't see: internal system names, topology, data flows. Before anything is generated:
- Write the fact graph (nodes, edges, labels) in plain text.
- Replace every internal name with its public role (
内部订单库 → orders DB, knlb-xxx → API gateway). If a node's only honest name is an internal one, the diagram is not publishable — generalize the whole example or pick a synthetic system.
- Drop what the viewer question doesn't need. Each removed node is attack surface removed.
- For "example/teaching" diagrams, prefer a synthetic reference system (a generic RAG app, a URL shortener) over any real deployment.
1. Freeze the contract
- Viewer question + one takeaway (what should the reader understand in 10 seconds?)
- Fact graph: node list, edge list with direction, exact label whitelist
- Grammar: layered architecture / request flow / deployment / state machine — one per figure
- Surface: doc inline (16:9 or 4:3), slide, or README hero
- Style — freeze it too. Default to the editorial language: warm-white canvas (#FAFAF7), navy (#1E3A5F) structure, accent colors only where they carry meaning, bold title + one-line takeaway caption, text-only cards (no provider logos/icons). A dark "dashboard chrome" theme (saturated neon boxes on navy) reads as a themed Mermaid export — "looks like Mermaid" is a reject reason even when every label is correct.
2. Choose the production branch
| Branch |
When |
Trade-off |
| T1 full generated (labels in image) |
short labels (≤4 chars each), ≤10 nodes, Chinese OK |
beautiful, but label fidelity must be audited |
| Hybrid (no-text generated plate + deterministic text overlay) |
long labels, many nodes, exact wording required |
labels guaranteed; two-step work |
| Retreat to deterministic (Mermaid/SVG/D2) |
topology must be exact, or two audits already failed |
always correct, less pretty |
Never invent nodes, edges, arrows, or numbers beyond the frozen fact graph — and don't let the model do it either.
3. Generate with the fact graph in the prompt
Structure the prompt as: scene grammar → node inventory with exact labels → edges with direction → style (flat vector, dark/light plate) → constraints ("every label must be exactly as listed, no extra text").
anycap image generate --model <model> --prompt "..." -o candidate.png
For Chinese-labeled T1 candidates, models with strong short-label rendering (e.g. nano-banana class) beat general image models.
4. Audit with image-read — labels, not vibes
anycap actions image-read --file candidate.png \
--instruction "List every text label in this diagram exactly as written, and every
arrow with its direction. Then: (1) any label not in this whitelist: <list>;
(2) any missing/duplicated label; (3) any arrow pointing the wrong way;
(4) any pseudo-text. Answer AUDIT-PASS or AUDIT-FAIL - <details>."
Typical failure modes: misspelled labels, merged nodes, reversed arrows, hallucinated extra services, pseudo-text watermarks. One concrete failure → regenerate with the failure named in the prompt. Two failed audits → switch branch (hybrid or deterministic). Don't keep rolling the same branch.
Measured ceiling (two shipped projects): T1 full generation with inline labels is reliable up to ~6 nodes / ~6 edges. At 8 nodes × 8 arrows, every roll drifted exactly one arrow — a different one each time (fix-one-break-another carousel). Layout-first prompt fixes (place a node so its required arrow is short and vertical) are a valid second-round move, but they don't move the ceiling. Retreating to a deterministic SVG/Mermaid for the densest view is a branch decision, not a failure — the deterministic source doubles as the editable truth layer.
5. Hybrid finishing when labels must be perfect
Generate a no-text plate (gate it CLEAN like any brand mark), then overlay labels deterministically (HTML/CSS, SVG, or an editor). The plate carries the mood; the overlay carries the truth.
6. Completion gate
- Every label matches the whitelist exactly; every edge direction verified
- No invented components; no internal names anywhere
- Readable at the consumer's real size (inline width, not 100% zoom)
- Provenance sidecar: fact graph, model, audit results, branch decisions
Anti-patterns
- "Draw my company's architecture" as a prompt — semantic leak by design
- Accepting a diagram because it "looks professional" while two arrows point backwards
- More than ~12 nodes in one generated figure — split into views
- Shipping the raster without keeping the fact graph — the next edit restarts from zero
- Shipping a single overview when the system has distinct questions — a small-multiples view set (overview / request flow / data pipeline / deployment), all projecting one frozen fact model, beats one overloaded poster. Views share node names, palette, typography, and connector style; change the model in one view and you must change all.
1---2name: anycap-architecture-diagrams3description: Produce architecture and system diagrams with AnyCap — fact-graph-first prompting, label fidelity audit via image-read, hybrid plates with deterministic labels, and a public-safety pass that strips internal system details before generation. Use when turning a system design into a shareable architecture figure with AnyCap image models.4---56# AnyCap Architecture Diagrams78Turn a system design into a clean architecture figure with AnyCap. Distilled from repeated real use; the loop below is the one that converges.910## 0. Sanitize the fact graph FIRST (public-safety pass)1112An architecture diagram is a浓缩的系统地图 — it leaks exactly what scans can't see: internal system names, topology, data flows. Before anything is generated:1314- Write the fact graph (nodes, edges, labels) in plain text.15- Replace every internal name with its **public role** (`内部订单库` → `orders DB`, `knlb-xxx` → `API gateway`). If a node's only honest name is an internal one, the diagram is not publishable — generalize the whole example or pick a synthetic system.16- Drop what the viewer question doesn't need. Each removed node is attack surface removed.17- For "example/teaching" diagrams, prefer a **synthetic reference system** (a generic RAG app, a URL shortener) over any real deployment.1819## 1. Freeze the contract2021- Viewer question + one takeaway (what should the reader understand in 10 seconds?)22- Fact graph: node list, edge list with direction, **exact label whitelist**23- Grammar: layered architecture / request flow / deployment / state machine — one per figure24- Surface: doc inline (16:9 or 4:3), slide, or README hero25- **Style — freeze it too.** Default to the editorial language: warm-white canvas (#FAFAF7), navy (#1E3A5F) structure, accent colors only where they carry meaning, bold title + one-line takeaway caption, text-only cards (no provider logos/icons). A dark "dashboard chrome" theme (saturated neon boxes on navy) reads as a themed Mermaid export — "looks like Mermaid" is a reject reason even when every label is correct.2627## 2. Choose the production branch2829| Branch | When | Trade-off |30|---|---|---|31| **T1 full generated** (labels in image) | short labels (≤4 chars each), ≤10 nodes, Chinese OK | beautiful, but label fidelity must be audited |32| **Hybrid** (no-text generated plate + deterministic text overlay) | long labels, many nodes, exact wording required | labels guaranteed; two-step work |33| **Retreat to deterministic** (Mermaid/SVG/D2) | topology must be exact, or two audits already failed | always correct, less pretty |3435Never invent nodes, edges, arrows, or numbers beyond the frozen fact graph — and don't let the model do it either.3637## 3. Generate with the fact graph in the prompt3839Structure the prompt as: scene grammar → node inventory with exact labels → edges with direction → style (flat vector, dark/light plate) → constraints ("every label must be exactly as listed, no extra text").4041```42anycap image generate --model <model> --prompt "..." -o candidate.png43```4445For Chinese-labeled T1 candidates, models with strong short-label rendering (e.g. nano-banana class) beat general image models.4647## 4. Audit with image-read — labels, not vibes4849```50anycap actions image-read --file candidate.png \51 --instruction "List every text label in this diagram exactly as written, and every52 arrow with its direction. Then: (1) any label not in this whitelist: <list>;53 (2) any missing/duplicated label; (3) any arrow pointing the wrong way;54 (4) any pseudo-text. Answer AUDIT-PASS or AUDIT-FAIL - <details>."55```5657Typical failure modes: misspelled labels, merged nodes, reversed arrows, hallucinated extra services, pseudo-text watermarks. **One concrete failure → regenerate with the failure named in the prompt. Two failed audits → switch branch (hybrid or deterministic).** Don't keep rolling the same branch.5859Measured ceiling (two shipped projects): T1 full generation with inline labels is reliable up to **~6 nodes / ~6 edges**. At 8 nodes × 8 arrows, every roll drifted exactly one arrow — a different one each time (fix-one-break-another carousel). Layout-first prompt fixes (place a node so its required arrow is short and vertical) are a valid second-round move, but they don't move the ceiling. Retreating to a deterministic SVG/Mermaid for the densest view is a branch decision, not a failure — the deterministic source doubles as the editable truth layer.6061## 5. Hybrid finishing when labels must be perfect6263Generate a no-text plate (gate it CLEAN like any brand mark), then overlay labels deterministically (HTML/CSS, SVG, or an editor). The plate carries the mood; the overlay carries the truth.6465## 6. Completion gate6667- Every label matches the whitelist exactly; every edge direction verified68- No invented components; no internal names anywhere69- Readable at the consumer's real size (inline width, not 100% zoom)70- Provenance sidecar: fact graph, model, audit results, branch decisions7172## Anti-patterns7374- "Draw my company's architecture" as a prompt — semantic leak by design75- Accepting a diagram because it "looks professional" while two arrows point backwards76- More than ~12 nodes in one generated figure — split into views77- Shipping the raster without keeping the fact graph — the next edit restarts from zero78- Shipping a single overview when the system has distinct questions — a small-multiples **view set** (overview / request flow / data pipeline / deployment), all projecting one frozen fact model, beats one overloaded poster. Views share node names, palette, typography, and connector style; change the model in one view and you must change all.
convergeai-labs