Generate explanatory diagrams and infographics that visually communicate concepts. Iterates autonomously until images are logically correct, text is clean, and the concept explanation is clear. Uses Nano Banana 2 (Gemini 3.1 Flash Image).
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Generate explanatory diagrams and infographics that visually communicate concepts. Iterates autonomously until images are logically correct, text is clean, and the concept explanation is clear. Uses Nano Banana 2 (Gemini 3.1 Flash Image).
automation
gated
allowed-tools
Bash, Read, Write, Glob, AskUserQuestion
user-invocable
true
metadata
{"version":"1.1","created":"2026-03-03T00:00:00.000Z","author":"Ruby","changelog":["1.1: Imported explanation principles from /animated-explainer, adapted for static images - truth gate before design, one worked example (prefer a failure), explainer-not-advertisement, mechanism/numbers separation with disclosed scale, locked no-crossing layout, nothing floats, visibly-unfinished parts, labels-first storyboarding; self-critique now checks facts and structure separately from looks","1.0: Initial version"]}
category
visual-communication
Create Explanatory Image
Purpose
Generate explanatory diagrams and infographics that clearly communicate a concept, iterating autonomously until the images are logically correct, text is accurate, and the visual explanation lands.
Core Principles
Adapted from /animated-explainer - what makes an image explain rather than decorate.
Apply them at decomposition time (Steps 2-3); the self-critique loop (Step 5) checks them.
The facts are the boss - interrogate before you design. Decide what's TRUE before
deciding what's compelling. If the image explains a real system, read its source of
truth (specs, architecture docs, the actual data) - not marketing copy or published
summaries. A concept the facts don't support dies at this step; that is the step
working, not failing. Never soften an unsupported claim into a diagram.
One concrete worked example beats any amount of abstraction. A diagram that traces
one real item end to end (this request, this invoice, this forecast) explains more
than an abstract schema. Prefer a failure case the mechanism catches - it shows the
machine doing its job and cannot be read as cherry-picking.
Explainer, not advertisement. Assume the viewer knows nothing. Open on the problem
the mechanism solves - the title states what's broken or the single insight, not the
product name. If the image only makes sense to someone who already understands the
system, it has failed.
Separate the mechanism from the numbers. Carry the mechanism with symbolic marks;
put real figures in ONE place (a footer stat, a single callout) or nowhere. If a mark
is not 1:1 with a thing, say so on the image (1 dot = ~30 runs). Never silently mix
scales. This also serves the 10-label limit.
Lock the layout - nothing crosses anything. One spine (a left-to-right flow or a
single loop), elements sitting ON it. Crossing arrows read as complexity, not as a
system. Prompt for it explicitly and reject renders where connections cross or wander.
Nothing floats - every element visibly comes from somewhere. The static cousin of
"nothing teleports": causality is carried by explicit arrows and numbered steps
(input -> transform -> output). An element with no incoming or outgoing connection is
decoration - cut it or wire it in.
Show the part that doesn't work. If part of the system is unproven, planned, or
broken, draw it dashed/hollow/greyed and label it. A diagram where everything works is
the least trustworthy thing you can draw - the unfinished part is the credibility.
Labels first. Write every text string before generating anything (Step 2 does
this) - if the labels alone, read in order, tell the story, the image will work. If
they don't, no amount of visual polish will save it.
Note complexity limits: 10 boxes max, 10 text labels max, 3 hierarchy levels max
Note known model limitations (misspellings on 4+ syllable words, font inconsistency)
Confirm output folder:
If user specified a folder, use it
Otherwise, create: [concept_slug]_diagrams/ in current directory
mkdir -p [folder]
Step 2: Decompose the Concept
Analyze the user's concept description and break it down:
Truth gate (Principle 1): If the image describes a real system or carries real
numbers, verify every claim against the source of truth (specs, docs, actual data) -
never from memory or marketing copy. Cut anything unsupported and tell the user what
was cut and why.
Core message: What single insight should the viewer walk away with? State it as
the problem or insight, not the product name (Principle 3).
Worked example (Principle 2): Can the concept be shown as one concrete item traced
end to end instead of an abstract schema? Prefer it; prefer a failure case the
mechanism catches.
Visual elements: List all shapes, icons, nodes, connections needed. Every element
must be wired into the flow (Principle 6) - an element with no incoming or outgoing
connection gets cut or connected.
Text labels: List every text string that will appear in the image - then read them
in order. If the labels alone tell the story, the image will work; if they don't, fix
the story before generating anything (Principle 8).
Honest state (Principle 7): Which parts are live vs planned/unproven/broken? Mark
the latter to render dashed, hollow, or greyed with a label.
This step is critical. The concept decomposition will almost always exceed Nano Banana's limits. Simplify ruthlessly:
Count visual elements - if > 10, merge or remove until <= 10
Count text labels - if > 10, shorten labels to 1-2 words, replace words with icons, or remove sub-labels
Move numbers out of the mechanism (Principle 4) - carry the mechanism with symbolic
marks; keep real figures in ONE place (a footer stat or single callout) or drop them.
If a mark is not 1:1, spend one label on the scale legend (1 dot = ~30 runs).
Check word complexity - replace any word with 4+ syllables with a simpler alternative:
Avoid (misspells)
Use Instead
SOVEREIGNTY
OWN SPACE, ISOLATE
ORCHESTRATION
COORDINATE, TEAMWORK
HIERARCHICAL
TOP-DOWN, LEADER
INFRASTRUCTURE
FOUNDATION, BUILD
OBSERVABILITY
MONITOR, WATCH
GOVERNANCE
CONTROL, AUDIT
PLAYBOOK
(usually OK but occasionally garbles)
Verify hierarchy - max 3 levels: title, main content, footer
Present the simplified plan:
Concept: [one sentence]
Layout: [type]
Elements: [count] / 10 max
Labels: [count] / 10 max
Complex words replaced: [list]
Cut as unsupported: [list, or none]
Unfinished parts (render dashed): [list, or none]
Step 4: Generate Initial Variants
Generate the specified number of variants (default 3), each taking a slightly different visual approach to the same concept.
For each variant:
Craft a narrative prompt following best practices:
Write descriptive paragraphs, not keyword lists
Include style tokens (background color, text color, font)
Specify layout positioning explicitly - prescribe element order and position
("left to right: A, then B, then C"); the model follows stated order far better
than inferred structure
Lock the layout (Principle 5): one spine (left-to-right flow or a single loop),
elements sitting on it, and say explicitly that no connection lines cross
For unproven/planned parts (Principle 7): request "drawn with a dashed outline,
unfilled" right next to that element's description, plus its label
Request "generous spacing" and "clean minimal design"
Generate using the Python script (handles JSON escaping properly):
Important: Use the Python script, not the bash script, to avoid JSON escaping issues with quotes in prompts.
Wait 2 seconds between API calls to avoid rate limits:
sleep 2
Step 5: Self-Critique Loop
For each generated image, run this analysis cycle. This is the core differentiator of this skill.
View the image: Use the Read tool on the PNG file to visually inspect it
Check for issues against this checklist. Looking at the render and checking the
facts are DIFFERENT jobs - layout bugs are invisible in the plan, factual bugs are
invisible in the render. Do both.
Looks (inspect the render):
Title text: Is it spelled correctly and readable?
All labels: Are they spelled correctly? Any garbled/mushy text?
Element count: Does the image have roughly the right number of elements?
Missing elements: Is anything from the concept missing entirely?
Duplicate elements: Are any labels or nodes repeated incorrectly?
Text placement: Are labels in the right positions relative to their elements?
Readability: Would this be readable on a mobile screen?
Structure (check against the Step 2 plan):
Logical structure: Does the layout match the requested concept? Are connections correct?
No crossings: Connection lines don't cross or wander (crossing reads as complexity, not system)
Nothing floats: Every element is wired into the flow with a visible connection
Honest state: Unproven/planned parts actually rendered dashed/hollow with their label
Truth (check against the Step 2 sources):
Factual accuracy: Every number and claim in the render matches the verified source; the scale legend is present if marks are not 1:1
Cold-viewer test: Would someone with zero context walk away with the core message? (Explainer, not advertisement)
Classify the image:
Good: No issues found, or only very minor ones. Keep as candidate.
Fixable: 1-2 specific issues that can be addressed by prompt adjustment. Iterate.
Redo: Fundamental layout/structure problems. Needs a different prompt approach.
For Fixable images, identify the specific fix:
Missing element โ Add explicit instruction for it
Misspelled word โ Replace with simpler word or remove
Wrong layout โ Be more explicit about positioning
Duplicate node โ Emphasize exact count ("exactly six nodes, not five, not seven")
Elements merged โ Use completely distinct words for each element
Crossing lines โ Restate the layout as an explicit ordered sequence with positions ("one horizontal spine, left to right: A, B, C; no lines cross")
Unfinished part rendered solid โ Repeat the dashed/hollow instruction immediately adjacent to that element's description
Floating element โ Name its connection explicitly ("an arrow from X to Y") or cut it
Regenerate with targeted fixes. Maximum 3 iteration rounds per variant.
Track the best version of each variant approach.
Step 6: Present Candidates
[APPROVAL GATE] - Present the best candidates to the user
Show each candidate image and describe:
What it got right
Any remaining minor imperfections
Which concept angle it takes
User options:
Pick a winner - Select one or more as final
Iterate on specific one - Request changes to a candidate
New direction - Describe a different visual approach
Combine elements - Mix aspects from multiple candidates
If user requests changes:
Apply modifications to the prompt
Regenerate and re-run self-critique (Step 5)
Return to this gate
Step 7: Finalize and Cleanup
Copy final images to the output folder with clean names:
