Generate the image-generation prompt for a `format: decision-tree` LinkedIn post — a single clean hand-drawn flowchart that routes the reader from a root question to a recommendation across 3 to 5 labelled branches, with the question rendered as the in-image title. Always square. Use when a decision-tree draft is approved, the user says "flowchart image", "decision tree image", "render the tree", or `post-cycle` reaches the visual step on a decision-tree draft. Saves the prompt to `concepts/<date>-<slug>/prompt.md`, updates the draft's `concept_path`, then calls OpenAI (`gpt-image-2`, if `OPENAI_API_KEY` is set) to render it to `images/<date>-<slug>/prompt.png`, and prints the prompt either way so the user can paste it into another image tool. Trigger phrases: "post-flowchart", "flowchart image", "decision tree 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 the image-generation prompt for a `format: decision-tree` LinkedIn post — a single clean hand-drawn flowchart that routes the reader from a root question to a recommendation across 3 to 5 labelled branches, with the question rendered as the in-image title. Always square. Use when a decision-tree draft is approved, the user says "flowchart image", "decision tree image", "render the tree", or `post-cycle` reaches the visual step on a decision-tree draft. Saves the prompt to `concepts/<date>-<slug>/prompt.md`, updates the draft's `concept_path`, then calls OpenAI (`gpt-image-2`, if `OPENAI_API_KEY` is set) to render it to `images/<date>-<slug>/prompt.png`, and prints the prompt either way so the user can paste it into another image tool. Trigger phrases: "post-flowchart", "flowchart image", "decision tree image".
post-flowchart
Build one ready-to-paste image-generation prompt that renders a LinkedIn decision-tree post as a single, legible flowchart: a root question at the top, 3 to 5 branches each labelled with a concrete condition, each landing on a named recommendation. It is the decision-tree sibling of post-image and post-carousel.
This is the one deliberate exception to post-image's "draw a metaphor, never a chart" rule. A decision tree is a diagram — the flowchart is the literal content, not a lazy stand-in. So this skill does not pick a metaphor or stage a subject in tension; it renders a clean, hand-drawn diagram that a reader can follow in seconds.
This skill writes a prompt to disk, updates the draft to link to it, then renders it with OpenAI's gpt-image-2 when OPENAI_API_KEY is set, saving the PNG into a gitignored images/ folder that mirrors concepts/. If the key is not set (or generation fails), the prompt is still saved and printed for the user to paste into an image tool by hand.
Use this only for format: decision-tree drafts. Text posts use post-image; carousels use post-carousel.
Office UI sync
This is the illustrator stage (the decision-tree branch of it) — emit end once concepts/<date>-<slug>/prompt.md is written, per office-emit-end. Use --stage illustrator --response-file concepts/<date>-<slug>/prompt.md. No separate stage; the file is named prompt.md so the dashboard finds it as it would any concept.
Inputs
Draft path like drafts/2026-06-16-which-vector-db-should-you-use.md. Required. The draft must be format: decision-tree with a ## Decision tree body (a Question: line plus 3 to 5 - If <condition> → <recommendation> branches and a Caption:), as written by post-writer's decision-tree path. If the draft is any other format, stop and point the user to post-image (text) or post-carousel (carousel).
Style flag as a token of args (sketch-on-white or hand-drawn). If absent, prompt the user; default to sketch-on-white. Inside /post-cycle, surface the choice. The photo style is not offered — a flowchart is line art with rendered labels, which the photo style refuses.
Size is always square, 1200 × 1200, ratio 1:1 — the LinkedIn-accepted general-feed size (see linkedin-image-specs). There is no size flag. A tall tree (4 to 5 branches) still lays out inside the square — keep it balanced and legible rather than reaching for more vertical room.
The skill does not accept posts/... paths.
Building the flowchart
Read the draft's ## Decision tree block and lay it out as a top-down flow:
Root: the Question: line, rendered as the in-image title at the top and as the single starting node of the flow.
Branches: one labelled edge per - If <condition> → <recommendation> line. The condition is the edge label; the recommendation is the leaf node it points to. Keep 3 to 5 branches — if the draft has more, stop and ask the writer to cut (the image must stay legible).
Leaves: each recommendation is a clearly drawn end node (a box with the named tool/approach).
Keep the layout clean and balanced: root at top, branches fanning down, leaves on one row or two. One level of nesting at most. No crossing edges, no clutter, generous whitespace. It must read in milliseconds at phone size.
Title overlay rules
Render the root question IN-image as the title (for both styles). Compress it hard — it is the biggest text on the canvas:
Trim to 3 to 7 words; strip quote marks, parentheses, em dashes, trailing punctuation.
Keep the core wording; drop filler ("should you", "do you have a").
All caps in the final overlay.
Legibility budget (hard rule — the image must read in one second)
This is the rule that decides whether the flowchart works. An AI image tool renders long text tiny and garbled, so you compress the draft's branch text before it goes in the image. Never render a full sentence in a node or on an arrow. The supporting detail (numbers, examples, the "why") lives in the caption only — it is already there; do not duplicate it into the diagram.
Hard caps on what is rendered in the image:
Condition label (on the arrow): at most 4 words. A noun phrase, not a sentence. No examples, no numbers, no colons, no parentheticals.
Recommendation node: at most 3 words. An imperative or a verdict.
Title: 3 to 7 words (above).
3 to 5 branches. If the draft has more, stop and ask the writer to cut.
Compress each branch to its essence. Worked example, from a real over-stuffed draft:
draft branch (too long for the image)
rendered label → node
If a fixed evaluator grades every run and a git revert undoes any damage → let it loop overnight (~100 runs, auto-revert)
GRADED + REVERSIBLE → LOOP OVERNIGHT
If the task reads untrusted input (Sentry errors, web pages, a stranger's repo) → never autonomous
READS UNTRUSTED INPUT → NEVER
If the blast radius is irreversible (prod data, live credentials, real money) → keep your hand on it
IRREVERSIBLE BLAST RADIUS → STAY HANDS-ON
Otherwise you are the only evaluator → supervise
NO EVALUATOR YET → SUPERVISE
If you cannot get a condition under 4 words without losing its meaning, the branch is doing too much — that nuance belongs in the caption, and the label keeps only the trigger. Big, sparse text beats complete text every time.
Styles
Two styles: sketch-on-white (default) and hand-drawn. Same diagram either way; only the rendering descriptors change. Include the chosen style's style spine and the tailored negative prompt block verbatim.
sketch-on-white
Style spine:
Minimalist hand-drawn flowchart: confident single-weight black ink lines on a solid
pure-white background, no color or fill, like a black marker on paper. Clean nodes
(rounded boxes) and labelled connector arrows, generous negative space, balanced top-down
layout. Reads in milliseconds. All text hand-lettered in black. Title hand-lettered in
black across the top.
hand-drawn
Style spine:
Hand-drawn flowchart in a warm 1960s mid-century-modern style: clean confident lines with
a lightly textured hand-sketched feel on warm off-white paper, restrained muted palette
(mustard, teal, burnt orange, ochre, cream) used sparingly, subtle paper grain. Rounded
nodes and labelled connector arrows, balanced top-down layout, generous negative space.
Reads in milliseconds. All text hand-lettered. Title hand-lettered across the top.
Diagram + label rendering block (append, per chosen style's lettering)
Render the root question IN-IMAGE as a big hand-lettered title across the top, large and
high-contrast. Below it draw a clean top-down flowchart: one starting node, 3 to 5 arrows
fanning down, each arrow labelled with a SHORT condition of at most four words, each arrow
ending in a clearly drawn node holding a recommendation of at most three words. All text
large and legible at phone size — few words, big letters. No full sentences, no small
print, no crossing arrows, no clutter, generous whitespace.
Negative prompt block (tailored — a flowchart IS allowed here)
Use this in place of post-image's block, because a decision flowchart is the literal subject and must not be suppressed:
Avoid: photorealism, 3D, gradients, halftone, watercolor, anime, emoji, blurry or garbled
text, crossing or tangled arrows, busy backgrounds, clutter, more than five branches. Keep
the layout clean and balanced and the text crisp and readable. No real or trademarked
characters, logos, or brands; all marks original. (A clean flowchart with nodes and arrows
is the intended subject — do not avoid it.) No cliché brains, lightbulbs, gears,
locks-and-keys, robot faces, magnifying glasses, chess pieces, rockets, plain handshakes,
or dollar signs.
Aspect ratio suffix (append last)
Aspect ratio 1:1, 1200x1200, square; title across the top, the flowchart filling the rest.
Persistence
Concept folder
For a draft at drafts/<YYYY-MM-DD>-<slug>.md, write to:
concepts/<YYYY-MM-DD>-<slug>/prompt.md
Create the folder if missing; overwrite prompt.md if it exists. The file must be named exactly prompt.md — the dashboard and scraper key off that name. concept_path points at it.
Keep hook_overlay set to the compressed root question so the dashboard can title the card. The scraper may later append post_url/post_path; do not author those by hand.
Prompt file body
The body is the full prompt as printed below: the chosen style's spine, the flowchart description (root question, each labelled branch and its leaf, in order), the diagram+label rendering block, the aspect suffix, and the tailored negative prompt block.
Draft linkage
Update the draft frontmatter concept_path: concepts/<YYYY-MM-DD>-<slug>/prompt.md via Edit (preserve the rest). Overwrite an existing value.
Image generation
Once prompt.md is written, run:
bun run generate-image concepts/<YYYY-MM-DD>-<slug>
This calls OpenAI's gpt-image-2 with the saved prompt, requesting a standard square image, and writes the finished PNG to images/<YYYY-MM-DD>-<slug>/prompt.png. No cropping or resizing — the model is only ever asked for square.
If OPENAI_API_KEY is not set, the command exits with OPENAI_API_KEY is not set — skipped. Expected, not a failure — record skipped and move on.
If it fails for any other reason, surface the one-line error and record failed. Never stop the skill or fail the run over a generation error — the prompt is already saved.
Only report generated when the command printed saved images/....
Output format
Style: <sketch-on-white | hand-drawn>
Size: square — 1200 x 1200
Root question: <ROOT QUESTION IN ALL CAPS>
Branches: <N>
Saved to: concepts/<YYYY-MM-DD>-<slug>/prompt.md
Image: images/<YYYY-MM-DD>-<slug>/prompt.png (generated) | skipped — OPENAI_API_KEY not set | failed — <one-line reason>
Linked from: drafts/<YYYY-MM-DD>-<slug>.md
Prompt:
<chosen style's style spine>
Flowchart: <root question as the start node, then each "condition -> recommendation"
branch listed in order>
<diagram + label rendering block>
<aspect ratio suffix>
<tailored negative prompt block>
No preamble. Just print so the user can paste.
Workflow
Read the draft. Confirm format: decision-tree; otherwise route to post-image/post-carousel.
Parse the root question and the 3 to 5 branches. If more than 5, stop and ask the writer to cut.
Resolve the style (default sketch-on-white; surface inside /post-cycle). Size is always square.
Compress to the legibility budget: title 3 to 7 words, each condition ≤ 4 words, each recommendation ≤ 3 words, all caps. Push every number, example, and qualifier into the caption, not the diagram.
Assemble the prompt in the exact output format.
Write concepts/<date>-<slug>/prompt.md.
Edit the draft to set concept_path.
Run bun run generate-image concepts/<date>-<slug> (see Image generation above). Record generated, skipped, or failed.
Print the prompt block, with the Image: line reflecting step 8.
Hard rules
Only for format: decision-tree drafts.
The concept file is always named prompt.md and concept_path points at it.
3 to 5 branches maximum; the image must read in one second.
Compress every rendered label to the budget: condition ≤ 4 words, recommendation ≤ 3 words, title 3 to 7 words. Never render a full sentence; the detail stays in the caption. Big sparse text over complete text.
Include the chosen style's style spine and the tailored negative prompt block verbatim. Never offer photo.
A clean flowchart is the intended subject — this is the explicit exception to the no-chart rule. Still forbid the other banned clichés and any real/trademarked logo.
Only report the image as generated when bun run generate-image actually printed saved images/... — a missing key or a failed call is skipped/failed, never silently reported as success.
A skipped or failed generation is never a reason to stop the skill — the prompt is already saved and usable on its own.
When NOT to use
The draft is a text post (format: text/absent) → post-image. Or a carousel (format: carousel) → post-carousel.
The user wants a style outside the two offered. photo is not offered for flowcharts.
Concept art for an already-published post in posts/. Out of scope.
