Post banners

Four category banners for articles, shorts and LinkedIn posts.

Delivery size 1200x675 (16:9). The generator caps at 836x470, so the shipped files are Lanczos upscales with mild unsharp, measured crisper on dense text than a plain resize. The untouched 836x470 originals are kept alongside as *-836.png for any future re-render, and the 4-up sheet as _source-sheet.png.

A 4px inset is taken off every edge before scaling: splitting the 4-up sheet leaves a bright 2-4px seam on whichever edges faced the interior divider, which would otherwise render as a stray light line along the banner.

All four are cleared for use as of the 2026-08-23 revision. Every one passed the checklist at the bottom of this file.

What these are for

Fallbacks, not a house style. A post that already has its own graphic keeps it. These exist so a post without one still has a proper header instead of nothing, and so the category is legible at a glance.

Never replace existing article art with a banner. As of 2026-08-23, 30 of 36 articles carry their own image and were left untouched; the 5 that had none were assigned a banner by category, and articles/index.md is a listing page and does not need one.

Usage

In an article’s front matter:

image: /assets/brand/banners/concept-explainer.png          # OG / social card
featured_image: /assets/brand/banners/concept-explainer.png # rendered at top of article

image drives the Open Graph and Twitter card. featured_image renders the banner in the article body. Set both to the same file unless there is a reason not to.

The set

architecture.png and research-technical.png both name the seven layers and both match /stack/. The other two carry no layer text.

architecture.png now carries a one-line gloss per layer, so it works as a standalone explainer rather than only as decoration. research-technical.png nests Tools, Memory, Retrieval and Systems inside the L2 Harness box and adds Identity & Scope under L3 Agent, which is the correct containment: those are harness components, not layers.

What the revision fixed

The first cut of architecture.png labelled the stack with a superseded draft taxonomy, printed eight labels for seven layers, and repeated “RETRIEVAL” across two different icons. research-technical.png had the same taxonomy problem plus rendering artifacts: “CONTEXT” and “MEMORY” each printed twice, and a corrupted OUTPU 1T'f above “OUTPUT”.

Both are resolved. research-technical.png also improved conceptually: Memory, Retrieval, Tools and IO/Systems are now drawn as components in a supporting row rather than as competing top-level layers, which is both correct and a clearer diagram than the original.

Known, accepted

• architecture.png draws eight plates for seven layers. Measured, not eyeballed: seven leader lines against eight plate rims, and the lines do not land consistently on a plate. The seven layers themselves are enumerated correctly and completely in the label column, which is what a reader actually reads, so the error is confined to the decorative stack. It is still wrong in a diagram whose entire point is seven, and the prior revision had this right with seven plates each carrying its own icon. Fix on the next pass by specifying “exactly seven plates, one leader line per plate, each line terminating on its own plate.”
• Heading tint on architecture.png now runs as a deliberate cool-to-warm progression (saturation 0.18 at L1 rising to 0.90 at L7) rather than the three arbitrary tinted rows of the previous revision. Resolved.
• Resolution is handled by upscaling rather than at source, since the generator caps at 836x470. Files are 1200x675 and will still be slightly soft against native 1200px artwork. Good enough for LinkedIn and Open Graph, both of which accept 16:9.

Checklist for any future banner that names the layers

The rule in mode.md: anything that names or depicts the layers must agree with /stack/. Before replacing a file here:

• Count the layers. Seven, not six, not eight.
• Read every label. Seven distinct names, in the documented order, each appearing once: Model, Harness, Agent, Context, Governance, Orchestration, Product.
• Check each icon is used once and is paired consistently wherever it repeats.
• Zoom to at least 300% and read every other string, including small text inside boxes, for garbling and duplication.

Human-Curated, AI-Enabled (HCAE)