Generate Visual Substack Article (Visual Agent — multi-image variant)

/generate-visual's sibling for long-form content. Same provider-free, prompt-only deliverable, but a Substack article warrants more than one image (a feed post never does): exactly one header/hero image every article gets regardless of body content (added 2026-09-14 per direct user feedback — Substack's editor has a dedicated header-image slot, and every published article should fill it), plus one brief per [IMAGE: <label>] marker the writer left in the body.

Arguments

• draft-id (optional) — generate visuals for one article. If omitted, run for every type: substack-article note with no linked visual note yet (a header image alone is enough reason to run this — in-body markers are additive, not a prerequisite).

Process

1. The header/hero image (mandatory, every article)

Read the article's title, subtitle, and overall thesis (not just the opening paragraph) and design one strong lead image that represents the piece as a whole — this is what shows in the Substack feed/email preview and at the top of the published post, so it needs to work as a standalone visual, not a diagram of one specific section's content. Prefer a concept that captures the article's central tension or claim over a generic title-card treatment.

2. Find the in-body marked spots

Read the full article body and collect every [IMAGE: <label>] marker the writer left. It's fine for an article to have zero markers — the header image alone still applies — don't invent a marker that isn't there.

4. Decide, per marker, what the image actually needs to show

Read the surrounding section's actual content, not just the label — the image must support that specific section's point (a mechanism diagram for a technical section, a comparison table/graphic for a vs.-style section, a simple chart if the section leans on real numbers). Don't default every marker to the same visual format — vary it per what that section actually needs, and check Content-Learnings/content-index.md's visual_format/image_concept columns so this article's own images don't repeat each other either. The header image (step 1) and every marker image should also look distinct from one another — the header sells the piece as a whole, a marker image explains one specific point, and they shouldn't read as the same graphic twice.

5. Research visual/design trends, when it would help

Same judgment-call gate as /generate-visual step 3 — skip when the header's or a marker's concept is simple enough that trend research wouldn't change the outcome.

6. Write one Visual-Brief-Note covering all of this article's images

Copy _Templates/Visual-Brief-Note.md into Visuals/ as YYYY-MM-DD--kebab-slug.md, platform: substack-article, draft_id. In the body, write the header image's Brief/Trend-Research/Prompt/ Why-this-visual block first (labelled ## Header Image), then repeat the same block shape once per in-body marker (labelled by which section it belongs to, e.g. ## Section Image: <label>) — so one note covers the whole article's image set (header plus however many markers exist, including zero) instead of scattering several near-identical files. For each prompt: same full spec as /generate-visual (subject, composition, hierarchy, lighting, mood, color, typography, aspect ratio). Substack's header-image slot and its in-body images both typically work well at a wider landscape ratio (e.g. 1456×816 or similar) than a social feed image; verify current guidance rather than assuming.

7. Link it back

Add the one Visual-Brief-Note's id to the article's visual_ids[] (a single id covering the header plus every marker image, not one id per image).

Report back

Show the header image's finished prompt first, then every section label alongside its own finished prompt — all clearly marked as prompts (not rendered images) ready to paste into ChatGPT Images one at a time. The header image goes in Substack's header-image slot when publishing; each marker's image gets placed manually at that marker's spot in the body once generated.

Hard rules

• Never skip the header image — every article gets one, regardless of whether it has any in-body markers.
• Never claim an image was generated when only a prompt was written.
• Never invent a marker that isn't actually in the article body.
• Never produce an in-body visual for a purely decorative purpose (the header image is exempt from this — its job is representing the whole piece, which is a legitimate purpose on its own).
• Never reuse another article's (or another marker's) prompt wording — each is written fresh, including the header image.