blog-sketches

name	blog-sketches
description	Create or refine optional blog post sketch images for carteakey.dev. Use this when a post needs a unique monochrome transparent illustration, when existing sketch assets need cleanup for feed/post rendering, or when wiring a post to the site-wide `image` / `imageAlt` workflow.

Blog Sketches

Create optional per-post sketch images that work as:

• a thumbnail in /blog/, /feed/, and homepage featured slots
• a lead image on the post page
• the post's OG/Twitter/JSON-LD image

Treat these images as editorial flavor, not hero art. They should feel like a small stamped sketch that belongs to the post, not a reusable category badge.

Non-negotiables

• One image per post. Do not reuse category art across multiple posts.
• Transparent background preferred. Avoid white or off-white boxed backplates.
• Dark-mode safe. The site already uses dark:invert, so the drawing should be pure monochrome and hold up when inverted.
• Readable at small sizes. The same asset needs to survive around 80px-96px in feeds and much larger on the post page.
• Optional means optional. If the image is weak, skip it.

House Style

• Monotone black-ink or pencil sketch
• Notion-ish / editorial / notebook-adjacent, but not cute clipart
• Clear central silhouette with 1 main subject and 2-5 supporting objects
• Square-ish composition preferred
• Moderate detail, not noisy crosshatching everywhere
• Bold line weight: Avoid flimsy, hairline digital strokes. Outlines must be strong, thick, and well-defined (similar to a 4B pencil or felt ink pen) to ensure legibility at 80px feed sizes.
• Organic sketch texture: Favor charcoal, pencil, or ink textures over clean, flat vector blocks or clipart shapes.
• Technical or personal scene tied to the specific post, not the tag
• No color accents, gradients, or textured paper slabs
• No border frame, caption card, or fake printed card baked into the image
• Avoid wide empty margins; trim the asset so the subject reads quickly

Subject Selection

Read the post front matter plus opening section before generating anything. Pick a scene that reflects the post's actual angle:

• workflow post: laptop, terminals, notes, diagrams, desk objects
• local inference post: workstation, GPU tower, benchmark notes, tokens/meters
• homelab post: router, cables, NAS, Pi, rack shelf, dashboard
• notes / PKM post: notebook, vault cards, graphs, bookmarks, filing motifs
• reflective / essay post: fewer objects, stronger central symbol

Do not generate generic "AI art" symbols unless the post is explicitly about that symbol.

File Conventions

• Final assets live in src/static/img/blog-sketches/unique/
• Use slugged names like post-slug-stamp-trim.png
• Front matter should point at the trimmed transparent asset

Example:

image: /img/blog-sketches/unique/agent-ide-stamp-trim.png
imageAlt: Transparent monochrome sketch of an AI coding workspace

Workflow

1. Read the target post and decide whether an image is worth adding.
2. Generate one post-specific sketch with the raster image workflow/tooling.
3. If the generated asset has a removable flat chroma-key background, use the installed imagegen helper:

python3 "${CODEX_HOME:-$HOME/.codex}/skills/.system/imagegen/scripts/remove_chroma_key.py" \
  --input tmp/imagegen/agent-ide-keyed.png \
  --out src/static/img/blog-sketches/unique/agent-ide-stamp-trim.png \
  --auto-key border \
  --soft-matte \
  --transparent-threshold 12 \
  --opaque-threshold 220 \
  --despill \
  --force

4. Add image and imageAlt to exactly one post.
5. Verify the result on:
   • /blog/
   • /feed/
   • the post page
6. Run npm run build.

Batch Workflow

Use a batch pass when many posts lack thumbnails, but still treat every image as unique. The safest pattern is:

1. List posts missing image front matter, excluding the template:

for f in src/posts/**/*.md src/posts/*.md; do
  [ -f "$f" ] || continue
  [ "${f##*/}" = "1990-01-01-template.md" ] && continue
  if ! sed -n '1,/^---$/p' "$f" | rg -q '^image:'; then
    printf '%s\n' "$f"
  fi
done

2. Read each post's front matter and opening section before prompting.
3. Generate one image per post with the built-in image_gen tool. Use one prompt per post, not one generic prompt for the set.
4. Preserve a stable ordered mapping from generated source file to post slug. The built-in tool saves under $CODEX_HOME/generated_images/...; sort by modification time when generation order matters.
5. Copy or process the selected source into the repo. Never reference an image directly from $CODEX_HOME/generated_images.
6. Remove the chroma key and trim transparent edges into src/static/img/blog-sketches/unique/{slug}-stamp-trim.png.
7. Normalize any residual keyed tint to neutral grayscale. This matters because faint green or magenta wash can survive inside antialiased pencil shading and looks wrong after dark-mode inversion.
8. Validate each final PNG:
   • has an alpha channel
   • transparent corners are 0
   • the subject has plausible coverage, roughly 0.12-0.60
   • the contact sheet reads clearly at thumbnail size
9. Add image and imageAlt only to the matching post. Do not touch posts that already have a thumbnail unless replacing one intentionally.
10. Remove temporary keyed sources and contact sheets after final assets are in src/static/img/blog-sketches/unique/.
11. Run npm run build.
12. Serve _site locally and spot-check /blog/, /feed/, and representative updated post pages. If Browser/Playwright is unavailable, at minimum fetch the pages and HEAD the sketch image URLs.

Useful batch post-processing shape:

from pathlib import Path
from PIL import Image

def trim_to_square_alpha(path: Path, out: Path, pad: int = 28) -> None:
    im = Image.open(path).convert("RGBA")
    bbox = im.getchannel("A").getbbox()
    if not bbox:
        raise ValueError(f"no visible subject: {path}")
    left, top, right, bottom = bbox
    left = max(0, left - pad)
    top = max(0, top - pad)
    right = min(im.width, right + pad)
    bottom = min(im.height, bottom + pad)
    cropped = im.crop((left, top, right, bottom))
    size = max(cropped.width, cropped.height)
    canvas = Image.new("RGBA", (size, size), (0, 0, 0, 0))
    canvas.paste(cropped, ((size - cropped.width) // 2, (size - cropped.height) // 2), cropped)
    canvas.save(out, optimize=True)

def neutralize_to_grayscale(path: Path) -> None:
    im = Image.open(path).convert("RGBA")
    pixels = im.load()
    for y in range(im.height):
        for x in range(im.width):
            r, g, b, a = pixels[x, y]
            if a == 0:
                pixels[x, y] = (0, 0, 0, 0)
                continue
            lum = int(0.299 * r + 0.587 * g + 0.114 * b)
            pixels[x, y] = (lum, lum, lum, a)
    im.save(path, optimize=True)

For front-matter edits in a batch, prefer a path-keyed script over a broad text replacement. Insert image / imageAlt after description when present, and skip files that already contain image:.

Prompting Guidance

Keep prompts concrete and post-specific. Ask for:

• transparent background when possible
• monochrome textured charcoal pencil / ink sketch
• bold hand-drawn outlines, thick pencil strokes, and strong shaded contours
• no clean vector lines, no flat digital clipart, no hairlines
• centered composition with a strong silhouette readable as a small blog thumbnail
• tangible real-world metaphors (e.g. massive open tome with tiny companion notepad for model offloading, or a hand stamping a checkmark on blocks for token verification) rather than generic abstract network circles
• no colored fills
• no paper card or white border
• no readable text, labels, letters, or numbers unless the post specifically needs them

Bad prompt shape:

• "make a sketch for AI agents"

Good prompt shape:

• "Transparent monochrome editorial sketch of a laptop workspace for an essay about searching for the perfect agent IDE: code editor on screen, floating app cards, notebook with flowchart, coffee mug, clean black-ink lines, no background card, square composition, readable as a small blog thumbnail."

For chroma-key batch generation, add a strict removable-background block:

Scene/backdrop: perfectly flat solid #00ff00 chroma-key background for
background removal.

Constraints: uniform #00ff00 background only, no shadows, no gradients, no
texture, no reflections, no floor plane, no readable text, no watermark, no
border, no card, no white paper backplate, and no #00ff00 inside the subject.

If the subject is green-heavy, use #ff00ff instead. Avoid blue keys for local inference, hardware, and UI sketches because screens, charts, and technical objects often pick up blue-toned shading.

Editing Existing Assets

If a generated image is close but still has:

• off-white paper background
• too much transparent padding
• weak separation from the site background

then key out the background, trim the transparent edges, and desaturate the remaining visible pixels to grayscale. The skill currently has no project-local scripts/prepare-sketch.mjs; use the installed imagegen chroma-key helper or add a local helper intentionally before referencing one.

Integration Notes

• The site templates already support optional image / imageAlt.
• Do not add images to many posts at once unless each one is unique.
• Keep archive/feed density intact; do not add large framed thumbnails to lists.
• If you need to place the image inside post content manually, use the existing Eleventy image shortcode rather than Markdown image syntax.
• Permanent thumbnail batches should update docs/CHANGELOG.md and versions.json according to the repo playbook.
• Treat unrelated worktree changes as user-owned. Do not clean or revert existing deleted temp/source files unless explicitly asked.