Nano Banana — Visual Abstract Generation

Create publication-quality scientific figures that use visual metaphors, physical analogies, and isometric depth to convey complex technical systems. These are the figures you see in Nature, Science, and Nature Methods — not boxes and arrows.

Visual Abstracts vs Standard Diagrams

	Standard diagram (`diagram` skill)	Visual abstract (this skill)
Prompt style	"API gateway connects to auth service"	"API gateway as a routing prism splitting request beams into wavelengths"
Visual output	Boxes, arrows, labels	Metaphors, depth, physical analogies, glow, transparency
Audience	Engineers reading architecture docs	Anyone — meaning is conveyed through visual metaphor
Use case	PRDs, architecture docs, ERDs	README heroes, paper figures, blog posts, talks

Use the diagram skill for standard technical documentation. Use this skill when you want a human to feel the system, not just read it.

When to Use

README hero images that explain a project at a glance
Paper figures for journals, conferences, preprints
Blog post graphics that convey architecture to non-engineers
Conference talk visuals
Any request for "rich", "expressive", "artistic", or "Nature-quality" diagrams

Quick Start

python3 ${CLAUDE_PLUGIN_ROOT}/skills/diagram/scripts/generate_diagram.py \
  "<your visual-abstract prompt>" \
  -o visual-abstract.png \
  --style visual-abstract \
  --doc-type journal \
  --resolution 2K

The --style visual-abstract sends dark-background, glow, and metaphor directives via system_instruction — separated from your content prompt. The --doc-type journal sets the quality threshold to 8.5/10. The generation engine iterates up to twice to meet this bar.

The Metaphor Translation Process

This is the core skill. For every technical concept, ask: "What does this LOOK LIKE physically?"

Step 1: Read the system

Explore the codebase or user description. Understand the architecture, data flow, and key abstractions. Use Read, Grep, and Glob to build a mental model.

Step 2: Identify 3-5 key concepts

Pick the concepts that are most important to communicate. Not everything needs to be in the figure — pick the ones that tell the story.

Step 3: Map each concept to a visual metaphor

Use the vocabulary table below, or invent a new metaphor. The principle: describe what the concept LOOKS LIKE physically, not what it IS abstractly.

Bad: "The data is compressed by the extraction module" Good: "Raw data flows as a thick, chaotic bundle of luminous fibers through a COMPRESSION FUNNEL — a physical narrowing that strips away noise. Discarded particles scatter and dim. What emerges is a thin, refined stream."

Step 4: Choose spatial layout

Layout	Best for	Example
Isometric exploded view	Layered systems (stack, pipeline)	recall architecture — hooks, pipeline, storage as vertical layers
Circular lifecycle	Cyclical processes	recall how-it-works — capture → store → recall → repeat
Cross-section	Internal structure	Database internals, compression pipeline
Constellation/network	Distributed systems	Microservices, mesh networks
Flow/river	Data pipelines	ETL, CI/CD, streaming

Step 5: Describe data flow

Not arrows. Describe the physical medium:

Luminous fibers — individual data streams
Particle flow — many small items moving together
Liquid/fluid — continuous data streams merging and splitting
Light beams — request routing, API calls
Electrical current — signal transmission, event propagation

Step 6: Add composition details

Colors, lighting, glow, depth, transparency. Add quantitative labels where meaningful (sizes, percentages, counts, durations).

Step 7: Append the style suffix

Always end the prompt with:

Style: Publication-quality scientific figure. Dark background (#0d1117). Isometric depth. Subtle glow effects on active elements. Clean sans-serif typography. No cartoon elements. Information density of a Nature figure.

Metaphor Vocabulary

Reference table for translating technical concepts into visual metaphors:

Concept	Visual metaphor	Why it works
Data compression	Funnel filtering particles, narrowing pipe	Physical narrowing = intuitive reduction
Storage layers	Geological strata, sedimentary layers	Time-layered accumulation, depth = age
Temporal decay	Brightness gradient (new = bright, old = dim)	Natural aging, fading memory
Protection / safety	Shield, vault, capsule, force field	Physical barrier = data protection
Data flow	Luminous fibers, pipes, streams, rivers	Physical medium carrying information
Filtering / selection	Sieve, prism splitting light, membrane	Physical separation of wanted from unwanted
API gateway	Routing prism splitting request beams	Light splitting = request routing
Load balancer	Distribution manifold, river delta	Physical flow splitting evenly
Cache	Fast-access crystal buffer, mirror surface	Crystalline = fast, ordered retrieval
Database	Deep storage matrix, geological core, vault	Depth = persistence, permanence
Network nodes	Constellation of connected stars	Natural clustering, visible connections
Context window	Layered transparent workspace, fish tank	Bounded space with visible contents
Summarization	Crystallization of raw material	Refinement from chaos to structure
Error / failure	Cracks, fractures, heat signatures, red glow	Physical damage = system damage
Budget / quota	Depleting gauge, meter, sand timer	Physical resource being consumed
Async operation	Detached/floating element, thin tether	Physical independence, loose coupling
Sync operation	Rigid connection, locked coupling, rail	Physical constraint, forced sequencing
Encryption	Sealed container, opacity, lock mechanism	Hidden contents = encrypted data
Queue	Pipeline with items waiting, conveyor belt	Physical ordering, first-in-first-out
Webhook / event	Spark, signal flare, ripple on surface	Sudden trigger, propagation

Composition Rules

Dark background (#0d1117) — maximum contrast, enables glow effects
Isometric perspective — creates depth without vanishing-point complexity
Information flows clockwise or top-to-bottom
Color semantics:
- Blue (#4a9eff) — active, primary, processing
- Green (#4aef7a) — storage, success, growth
- Amber (#ffb347) — recall, retrieval, attention
- Orange (#ff6b35) — protection, warning, critical path
- Red (#ff4444) — error, failure, danger
- Cyan (#00d4aa) — data pipeline, transformation
- Gray (#666) — dormant, inactive, deprecated
Glow on active/current elements, dim on dormant/old
Labels integrated into the visual — not floating text boxes
Quantitative data where meaningful (sizes, percentages, counts)
No cartoon elements — scientific illustration aesthetic
Sans-serif typography — Geist, Inter, or Helvetica style
Version badge in top-right corner when applicable

Prompt Template

Use this structure for every visual abstract:

Create a publication-quality scientific figure. Title: '<title>'. Dark background (#0d1117).

<1-2 sentence system description — what the figure is about>

LAYOUT: <isometric exploded view | circular lifecycle | cross-section | constellation | flow>

<ELEMENT 1 — name> (<color tone>):
<Detailed visual metaphor description. What does it look like? How does it
behave? What physical phenomenon does it represent? Include quantitative
labels if meaningful.>

<ELEMENT 2 — name> (<color tone>):
<Same treatment.>

<ELEMENT N>:
<...>

<DATA FLOW description — how information moves between elements, described
as a physical medium>

BOTTOM: '<tagline or key specs>'

Style: Publication-quality scientific figure. Dark background (#0d1117).
Isometric depth. Subtle glow effects on active elements. Clean sans-serif
typography. No cartoon elements. Information density of a Nature figure.

Examples

Example 1: Memory System Lifecycle

Diagram skill would say:

Memory system with session capture, episodic storage, and context injection.
Stop hook captures data, SessionStart injects it.

Visual abstract prompt:

Create a publication-quality scientific figure. Title: 'recall v1.2.0 —
How It Works'. Dark background (#0d1117).

This figure shows the lifecycle of a memory system for an AI coding
assistant. Use rich visual metaphors — NOT boxes and arrows.

LAYOUT: Circular lifecycle flowing clockwise, isometric perspective with
depth. Three phases arranged as segments of a circle.

PHASE 1 — CAPTURE (top, blue #4a9eff tones):
Show a SESSION as a glowing terminal window that is closing/fading. From
it, streams of data flow outward like luminous particles through fiber
optic cables. These streams represent the raw transcript — show it as a
thick, chaotic bundle (3.4MB label). The bundle passes through a
COMPRESSION FUNNEL — a visual narrowing that strips away noise (show
discarded particles scattering away, dimming). What emerges is a thin,
refined stream (42KB label, 98.8% compression). This refined stream flows
into a HAIKU BRAIN — a small, elegant neural node that transforms the
stream into structured knowledge. Git commit icons as anchor points.

PHASE 2 — STORE (right side, green #4aef7a tones):
Show EPISODIC STORAGE as stacked layers of translucent markdown pages —
like geological strata. Newest layers on top are brighter. Older layers
dim with temporal decay. A SEMANTIC layer floats nearby as a crystalline
knowledge graph — currently dormant with a subtle pulsing glow. A WORKING
STATE capsule sits as a protected vault.

PHASE 3 — RECALL (bottom-left, warm amber #ffb347 tones):
A fresh terminal window lights up. From episodic storage, relevant memories
flow back as curated streams through a BUDGET GATE — a metered injection
point labeled '4K chars'. Inside the session, recalled memories integrate
into Claude's context window as layered translucent panels.

COMPACTION PROTECTION (center, orange #ff6b35):
A shield around the working state. When context compacts (depicted as a
crushing force), the shield preserves the working state while everything
else is swept away.

Style: Publication-quality scientific figure. Dark background. Isometric
depth. Subtle glow effects. No cartoon elements.

Example 2: CI/CD Pipeline

Diagram skill would say:

CI/CD pipeline with GitHub Actions, build, test, and deploy stages.

Visual abstract prompt:

Create a publication-quality scientific figure. Title: 'Continuous Delivery
Pipeline'. Dark background (#0d1117).

LAYOUT: Horizontal flow, left to right, with industrial/manufacturing
aesthetic. The pipeline is a physical processing facility.

SOURCE (far left, blue #4a9eff):
Code commits arrive as luminous data packets flowing through fiber optic
cables from a REPOSITORY — depicted as a crystalline archive with branching
structures (git branches as physical tree branches with glowing tips).
Multiple developer nodes feed into the repository like tributaries.

BUILD STAGE (cyan #00d4aa):
A FORGE — an industrial smelting chamber where raw code is compiled. Show
heat signatures and transformation. Dependencies flow in as raw materials
from a PACKAGE REGISTRY (shelved containers). The output is a refined
artifact — a glowing compiled binary or container image.

TEST STAGE (amber #ffb347):
A QUALITY CHAMBER with inspection beams scanning the artifact from multiple
angles. Show test suites as parallel analysis beams — unit tests as fine
beams, integration tests as broad sweeps, e2e tests as full-spectrum scans.
Failed tests produce red fracture lines. A QUALITY GATE at the exit only
opens when all beams pass (green).

DEPLOY STAGE (green #4aef7a):
The artifact enters a DISTRIBUTION MANIFOLD that splits the deployment
stream into environment channels — staging (dim), production (bright).
Show canary deployment as a thin test stream alongside the main flow.
Production servers depicted as active processing nodes in a constellation.

Style: Publication-quality scientific figure. Dark background. Industrial
manufacturing aesthetic with digital overlay. Subtle glow. No cartoon
elements.

Generation Options

Flag	Default	Purpose
`--doc-type journal`	Use this	8.5/10 quality threshold — highest standard
`--resolution 2K`	1K	Higher resolution for print/retina
`--input <path>`	—	Edit an existing visual abstract
`-v`	off	Verbose output (shows scores, critiques)
`--timeout 120`	120s	Increase for very complex prompts

Editing Visual Abstracts

To modify an existing visual abstract:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/diagram/scripts/generate_diagram.py \
  "Add a monitoring layer showing real-time metrics as oscilloscope traces" \
  --input existing-abstract.png \
  -o existing-abstract-v2.png \
  --style visual-abstract \
  --doc-type journal

Or use the command: /nano-banana:edit existing-abstract.png "Add monitoring traces"

Tips for Higher Scores

Keep text labels SHORT — AI image models hallucinate spelling in longer text. Use 1-3 word labels.
Be spatially explicit — "above", "below", "left of", "flowing into" are more effective than abstract relationships.
Include quantitative data — sizes, percentages, counts add information density and credibility.
Specify glow/dim for every element — don't leave visual states ambiguous.
More detail = better results — the 9.5-scoring prompt was ~1500 words. Don't be brief.
Name the physical metaphor explicitly — "a COMPRESSION FUNNEL" not just "compression."

Gotchas

Always pass --style visual-abstract: This sends dark-background and glow directives via system_instruction, cleanly separated from your content prompt. Without it, you get the default white-background technical style.
Spelling artifacts: AI-generated text in images often has minor character substitutions (e.g., "3.4MB" may render as "S.4MB"). Keep labels short and essential. The reviewer catches severe cases.
Iteration 2 is key: First iterations score 6-7.5 due to spelling/layout issues. The review-and-iterate loop typically pushes to 9+. Don't set --iterations 1.
Timeout: Complex prompts with many elements may need --timeout 180.

visual-abstract