| name | infographic |
| description | This skill should be used when the user asks to create an infographic from repository documentation, Markdown docs, ADRs, specs, diagrams, research summaries, runbooks, or inline context. It writes copy-paste-ready image prompts by default, and if the user explicitly asks to generate, render, make, or create the actual image and an image generation tool is available, it should pass the generated prompt to that tool. |
Infographic Prompt Writer And Generator
Create a structured image-generation prompt from repo documentation or inline
context. Return the prompt for handoff by default. If the user explicitly asks
for an actual generated image and an image generation tool is available, call
that tool with the structured prompt after building it.
Inputs
context (required): A repo Markdown path/link, @file: reference, or inline
text block. If it is a path or link, read the source before writing the prompt.
If it is missing or unreadable, ask for it.
orientation (optional): landscape by default. Also accept portrait or
square.
level (optional): lite by default. Also accept medium or heavy.
format (optional): markdown by default. Also accept json.
mode (optional): infer from the request. Use prompt by default. Use
generate only when the user explicitly asks for an image to be generated,
rendered, created, or made.
style_reference (optional): a design system, brand guide, style guide,
product screenshot, prior visual, or other repo artifact that should influence
the visual treatment.
reference_images (optional): one or more prior explainers, screenshots,
mockups, or style references to preserve, remix, or edit. If provided,
identify each image by index and role in the prompt.
Normalize common aliases:
wide, 16:9, deck, slide -> landscape
vertical, 9:16, poster, mobile -> portrait
1:1 -> square
md, sections, structured -> markdown
object, schema, strict-json -> json
render, make image, create image, actual image, generate image ->
generate
prompt only, handoff, copy-paste, do not generate -> prompt
Mode Selection
Default to prompt handoff. This keeps the skill useful in environments without
image tools and avoids surprising generation when the user only asked for a
brief.
Use mode=generate only when all of these are true:
- The user asks for an actual generated image, not just an image prompt.
- An image generation tool is available in the current tool set.
- The tool can accept a text prompt, or a text prompt plus the provided
reference images.
When mode=generate:
- Build the same structured prompt that would be returned in prompt mode.
- Include the selected
model, quality, and size inside the prompt or tool
call when the tool supports those parameters.
- Pass the prompt to the image generation tool.
- After tool completion, return a short note identifying the generated image
result and the key parameters used. Do not include a long duplicate prompt
unless the user asked to see it.
When mode=prompt, no image tool is called. Return the prompt and parameter
summary exactly as a handoff artifact.
If the user asks to generate an image but no image generation tool is available,
return the handoff prompt and say that generation requires an attached image
generation tool.
If the user explicitly asks not to generate, do not call an image generation
tool even if one is available.
Style Handling
Use explicit style references when provided. If none are provided, inspect only
nearby, obvious repo artifacts when available, such as:
DESIGN.md
STYLEGUIDE.md
BRAND.md
docs/design/
docs/brand/
- screenshots, mockups, or prior explainers mentioned by the user
If no style reference exists, use a neutral technical-documentation style:
- clean structured surfaces
- legible sans-serif typography
- restrained accent color chosen from the repo or source material when possible
- compact panels with crisp borders
- clear hierarchy and high contrast for labels
- no decorative clutter, generic stock-photo treatment, or unsupported logos
Treat style references as visual guidance only. Do not use them as evidence for
product behavior, architecture, maturity, security, privacy, timelines, or
implementation details.
Source Handling
When context points to repo documentation:
- Read the source and use its front matter status when present.
- Treat
accepted, approved, or equivalent final-status docs as
authoritative for their stated scope.
- Preserve
proposed, draft, experimental, and research uncertainty. Do
not make unresolved design choices look settled.
- If the source is under
research/, frame it as exploratory unless the prompt
context explicitly asks for a research infographic.
- Use canonical project terms from the source. If a glossary exists, check it
when labels or component names are ambiguous.
- Do not invent implementation details, production readiness, security or
privacy guarantees, timelines, metrics, owners, or dependencies.
When context is inline text:
- Use only the supplied claims unless the user explicitly asks you to inspect
repo files.
- Keep unknowns out of the image or label them as open questions.
Density Levels
Use the level parameter to choose how much information the image should carry.
| Level | Use For | Content Budget | Layout Guidance | Quality Guidance |
|---|
lite | quick explainers, meeting recaps, social/deck visuals | 1 headline, 3-5 key points, 3-6 labels | one dominant flow or hub-and-spoke, generous spacing | quality=medium by default; use high when text accuracy matters |
medium | architecture explainers, decision summaries, runbook overviews | 1 headline, 5-8 key points, 6-10 labels | 2-4 sections, clear arrows, compact callouts | quality=high for diagrams, slides, small labels, or stakeholder-facing output |
heavy | dense technical posters, review artifacts, system maps | 1 headline, 8-12 key points, 10-18 labels | multi-panel poster, swimlanes, layered system map, or timeline | quality=high because dense labels and layout precision need maximum fidelity |
For heavy, explicitly request high quality or maximum fidelity from the image
model because small labels and dense layouts need more precision.
Model Parameters
When returning a prompt handoff or preparing a tool call, include suggested
generation parameters unless the caller requested prompt text only:
model: default to gpt-image-2 for new image-generation workflows unless
the caller specifies another model.
quality: choose from the density table. Prefer high whenever the image has
dense labels, legends, axes, footnotes, small text, or stakeholder-facing
diagrams.
size: choose for the final display canvas rather than relying on later
cropping:
landscape: 1536x864 for deck slides, or 1680x944 for a slightly
roomier 16:9 explainer canvas.
portrait: 1056x1488 for poster/deep-dive explainers.
square: 1024x1024 for compact concept maps.
overview: 1536x1024 when the user asks for an architecture overview or
a wider-than-slide reasoning board.
Prompting Rules
Write the image prompt like an artifact spec, not a vague illustration request.
For examples of good patterns, failure modes, and repairs, read
references/prompt-patterns.md only when the request asks for examples, the
source is dense, or the first prompt draft feels underspecified.
Include these parts in a stable order:
- Deliverable and canvas: infographic, orientation, aspect ratio, intended
audience.
- Style reference: explicit source if provided; otherwise a short neutral style
summary.
- Main message: the one idea the viewer should understand first.
- Required content: exact labels, components, relationships, and callouts.
- Layout: panel structure, flow direction, placement, hierarchy, and spacing.
- Reference image handling: name each input image by index and role when
references are provided.
- Text rules: quote exact in-image text, keep labels short, require verbatim,
legible typography.
- Constraints: no invented facts, unsupported arrows, generic stock-photo
treatment, decorative clutter, illegible tiny text, or logos not present in
the source/style reference.
Keep the prompt concrete and skimmable with labeled sections. Prefer short lines
over a long paragraph.
If format=json, write the prompt as a valid JSON object. Use strings and
arrays, no comments, no trailing commas, and no Markdown inside string values.
Keep the JSON shallow enough for a human to inspect quickly.
For text in the image:
- Put exact visible labels in quotes.
- Ask the model to render quoted labels verbatim, with no extra characters,
duplicate text, paraphrasing, or spelling changes.
- Specify typography, size, color, and placement when label fidelity matters.
- For uncommon terms or tricky labels, repeat the exact label once and spell out
difficult parts letter-by-letter, for example
S-S-O Callback.
- Use title case only when it matches the desired label.
- Avoid paragraphs inside the image. Use short labels and callouts.
- For dense diagrams, ask for large readable labels and strong contrast.
For reference images and edits:
- Identify each input image by index and role, for example
Image 1: prior release-process explainer, Image 2: product screenshot.
- State how the images interact, for example
use Image 1's layout structure and Image 2's visual treatment.
- Separate what changes from what must remain invariant. Use
change only...
and keep everything else the same for surgical edits.
- Restate invariants on every follow-up iteration: canvas, layout, canonical
labels, source status, colors, typography, and component boundaries.
- For edits to existing artifacts, explicitly preserve saturation, contrast,
camera/viewpoint, layout, arrows, labels, surrounding objects, and brand/style
elements unless the requested change depends on modifying them.
For diagrams and architecture flows:
- Use arrows only where the source states a relationship or data/control flow.
- Show open questions as "Open Question" or "Pending Decision" panels rather
than implying an answer.
- Keep similarly named components distinct when the source distinguishes them.
- Preserve source-stated trust boundaries, lifecycle states, ownership, and
failure modes without expanding them.
Orientation Mapping
landscape: 16:9, best for decks, architecture maps, horizontal flows.
portrait: 9:16 or 4:5, best for poster-style summaries and vertical process
stacks.
square: 1:1, best for compact concept maps.
State both the orientation and aspect ratio in the prompt.
Iteration Guidance
For multi-step prompt refinement, start with a clean base prompt and make one
small change per follow-up. Re-specify critical invariants each time:
- canonical component names and quoted labels
- source-backed arrows and boundaries only
- source status and open questions
- style reference or neutral style summary
- target canvas and legibility constraints
If a generated image drifts, repair the drift directly instead of expanding the
prompt broadly. Example: restore the original API Service and Worker Queue placement; change only the footer wording.
Final Checklist
Before returning a prompt, check that it:
- includes suggested
model, quality, and size parameters
- uses exact canonical terms and quoted visible text
- preserves
draft/proposed/accepted or equivalent status correctly
- avoids invented topology, implementation details, guarantees, and timelines
- fits the requested density and canvas without relying on later cropping
- forbids illegible microtext, generic stock-photo treatment, decorative
technology art, and unsupported arrows
Output Format
For mode=prompt, return only the generated prompt and a short parameter
summary. Do not include research notes or implementation commentary.
For mode=generate, call the available image generation tool with the generated
prompt. After the tool call, return only a concise result note and the key
parameters used.
For copyable format=markdown and format=json examples, read
references/output-formats.md only when you need the exact output shape.
If the source is too large for a single useful image, summarize the most
important theme for the requested level and say which parts were excluded from
the prompt.