handdrawn-schematic

name

handdrawn-schematic

description

Generate a hand-drawn, whiteboard-sketch schematic that explains a concept, pipeline, architecture, algorithm, or mechanism at a glance, on a pure white background. Composes a friendly hand-lettered infographic prompt (wavy marker strokes, numbered panels flowing left to right, chunky chalk arrows, hand-written notes beside shapes) and by default renders it to a PNG via the bundled codex image_generation tool (ChatGPT OAuth); falls back to emitting the copy-paste prompt when codex is unavailable. Use when the user wants a hand-drawn / hand-written / whiteboard-style schematic, a sketch diagram that shows a schema, a friendly explainer figure, a doodle-style method overview, or 손그림/화이트보드 스타일 도식/개념도. For matplotlib data plots use scienceplot-py or xkcd-py; for a multi-style (editorial, blueprint, neon, poster) prompt-only composer use wide-slide-illustrator; for the paper-review method+results pair use journal-club-review.

Hand-Drawn Schematic

Turn any schema (a concept, pipeline, architecture, algorithm, or mechanism) into a single friendly hand-drawn whiteboard illustration on a pure white background: wavy marker strokes, numbered panels flowing left to right, chunky chalk-style arrows, and short hand-written notes drawn beside the shapes. Every label, number, and equation is lettered by the same hand into the one drawing, so it reads like a research postdoc's explainer sketch for a small-group talk.

By default this skill renders a PNG with the bundled codex image_generation tool (ChatGPT OAuth, no API key). When codex is not available it falls back to emitting the finished prompt so the user can paste it into any image generator.

When to use

• "Draw a hand-drawn schematic of how X works"
• "Make a whiteboard-sketch figure explaining this pipeline / architecture"
• "Sketch a doodle-style diagram of this algorithm"
• "이 개념/구조 손그림 도식으로 그려줘" / "화이트보드 스타일 개념도 만들어줘"

Do NOT use for:

• Real data plots (matplotlib, seaborn, plotly) → scienceplot-py (publication science+nature style) or xkcd-py (sketch-style matplotlib from data).
• A multi-style prompt-only composer across six looks (Editorial Magazine, Engineering Blueprint, Swiss Minimalist, Dark Tech/Neon, Scientific Poster) → wide-slide-illustrator.
• The paired method + results infographics inside a paper walkthrough → journal-club-review (this skill is the reusable single-figure generator extracted from it, retuned to a white background).

Inputs

Whatever the user gives you about the schema to draw. Gather (or infer from the provided source: a paragraph, a code file, a paper section, a research-log entry):

1. Headline — the figure title, <= 8 words.
2. Subtitle — one sentence saying what the figure shows.
3. Panels — the 3 to 6 stages/steps the schema breaks into, left to right, each with a short label and a concrete visual description (boxes, arrows, small charts, symbols, the exact equations/labels to hand-letter).

Do not hallucinate domain content. Pull equations, symbols, axis labels, and parameter names verbatim from the source when one is provided.

Workflow

1. Build the figure brief

Compose a small brief from the schema (schema and rules in references/style-block.md):

{
  "headline": "<= 8 words, the figure title",
  "subtitle": "one sentence describing what the figure shows",
  "panels": [
    {"name": "Panel label", "content": "visual description: boxes, arrows, labels, small charts"},
    ...
  ]
}

• Panels flow left to right as a pipeline / narrative (input -> ... -> output), or as parallel facets of one idea. Aim for 3-6; widen the most content-heavy panel slightly.
• content is a visual description (what to draw), not prose.
• Every label is a hand-written note drawn beside the shape, never a pasted box or floating caption.
• Panels are English-only and keep labels short (<= 6 words per line), even when the surrounding conversation is in another language.
• If a panel contains a chart (bars, curve, box plot), say so explicitly and demand it stay hand-drawn: "wobbly hand-drawn marker bars with uneven tops, hand-lettered numbers, NOT a clean digital/matplotlib chart." Otherwise the model renders a crisp vector chart that clashes with the sketch.

2. Compose the full prompt

Read references/style-block.md and concatenate, in order: the SINGLE COHESIVE block, the OVERALL LOOK block (pure white background), TYPOGRAPHY, COMPOSITION, MOOD, then the title bar + panels filled from the brief, then the English guard. The style wording is load-bearing; use it verbatim, do not paraphrase.

3. Render (default) or fall back to prompt-only

Check codex first:

codex login status        # need "Logged in using ChatGPT"

If logged in, write the composed prompt to a temp file and run codex, saving the PNG where the user wants it (default ./schematic.png):

mkdir -p <out_dir>/.codex-logs
codex exec \
  --sandbox workspace-write \
  --skip-git-repo-check \
  --cd <out_dir> \
  -o <out_dir>/.codex-logs/schematic.md \
  "Use the image_generation tool to create the following wide 18:9 infographic slide and save it as ./schematic.png in the current directory. Do not edit any other files. Report only the saved file path.

PROMPT:
<composed prompt>"

For several figures, launch the codex jobs in the same turn with run_in_background: true so they render concurrently (~1-2 min each). After a job finishes, verify the PNG exists and is non-empty before reporting it:

[ -s <out_dir>/schematic.png ] && echo ok

If codex is not logged in (or the user asked for "prompt only" / "no render"), skip generation and deliver the composed prompt inside a fenced code block so the user can paste it into their own image generator (ChatGPT Image, gpt-image, DALL-E, Midjourney). Note the aspect-ratio caveat: gpt-image-1 supports 1024×1024 / 1024×1536 / 1536×1024, so generate 18:9 at 1536×1024 and crop, or use a service with custom dimensions.

4. Deliver

• Tell the user the saved PNG path (or hand over the prompt if prompt-only).
• If a render came out garbled (broken labels, vector-looking chart), say so and offer one retry; text rendering usually needs 1-2 tries.

Notes

• Rendering depends on a logged-in bundled codex runtime (ChatGPT OAuth). No API key is used. Without it the skill still produces the finished prompt.
• The only surface difference from the journal-club friendly-whiteboard block is the background: pure white (#FFFFFF), no cream/beige tint. Everything else (palette, wavy strokes, numbered badges, chalk arrows) is unchanged.

Files

• references/style-block.md: the canonical pure-white style block (verbatim, load-bearing), the figure-brief schema, the English guard, and the exact codex exec invocation.