Prep Images for Embed

The one thing to remember: upload JPEG, not WebP. ImageKit converts to WebP for browsers that accept it and falls back to JPEG for those that don’t. Upload WebP and the fallback becomes a PNG up to 3× larger. Measured, not assumed — see Format.

When to use this skill

• The user drops screenshots and wants them in a changelog entry, splash page, spec, or any markdown surface
• “Add these screenshots”, “put these in the changelog”, “upload these”
• Local image references need converting to CDN URLs
• An image is destined to become an og:image
• Any markdown surface should become image-heavy without the user manually uploading and pasting links

Why this exists

Screenshots arrive as Screenshot 2026-08-14 at 9.11.31 PM.png — 400KB of PNG with a filename that is useless to a search engine and a reader. Getting one into a changelog properly means renaming it, resizing it, stripping metadata, hosting it somewhere with correct headers, and writing alt text. That’s six steps of tedium per image, which is why posts end up with no images at all.

This skill makes it one call.

Prior art worth knowing: content-farm/plugin-modules/image-gin already does ImageKit upload — but from inside Obsidian, via a modal, on a note you have open. This skill is the agent-callable path. The auth pattern here (Basic base64(privateKey + ':')) is lifted from image-gin’s imagekitService.ts deliberately, so the two don’t diverge.

Prerequisites

# ~/.secrets (chmod 600) — same pattern as generate-consistent-og-images
IMAGEKIT_PRIVATE_KEY="private_..."
IMAGEKIT_PUBLIC_KEY="public_..."
IMAGEKIT_URL_ENDPOINT="https://ik.imagekit.io/<id>"

Tooling: sips (macOS built-in) does resize and JPEG. exiftool strips metadata. cwebp is only needed for the discouraged --format webp path.

The call

node scripts/prep-images.mjs \
  --slug plane-first-deploy \
  --repo self-host-stack \
  --src "~/Desktop/Screenshot 2026-08-14 at 9.11.31 PM.png" \
  --name "Railway__Usage-By-Project" \
  --alt "Railway usage dashboard showing \$46.59 current usage broken out across six client projects" \
  --emit html

Batch by repeating --src / --name / --alt as ordered triples. Every source needs its own name and alt — the script refuses otherwise, on purpose.

The division of labour (this is the important part)

The script does the mechanical half. The agent does the judgement half, and the judgement half is what makes the output worth anything:

Read every image before naming it. You have vision; use it. A name derived from the file’s timestamp is worthless. A name derived from what is actually on screen — Railway__Usage-By-Project, Plane__Setup-Form, Postiz__Memory-Breakdown--TWF — makes the filename carry SEO weight and keeps the CDN folder browsable six months later.

Alt text is not optional

The script rejects alt text under 15 characters, and rejects anything opening with image, screenshot, photo, picture, graphic, untitled, tbd, or todo.

That is deliberate friction. alt="screenshot" is worse than no alt at all, because it looks finished — it satisfies a linter while telling a screen-reader user nothing and giving a search engine nothing.

The test: a reader who cannot see the image should learn the same thing a viewer would.

✅ alt="Railway usage dashboard showing $46.59 current usage broken out across
       six client projects, with the-water-foundation highest at $14.87"

❌ alt="Railway screenshot"
❌ alt="dashboard"
❌ alt="image of the usage page"

Where the image is decorative and genuinely adds nothing — a divider, a texture — it should carry alt="", not bad alt text. But then ask whether it belongs at all.

Format: upload JPEG, always

This is the finding that most changes behaviour, and it is measured on this account, not inferred from docs.

ImageKit content-negotiates on Accept. What you upload determines the fallback:

Uploading JPEG wins on both paths. ImageKit’s encoder beats hand-rolled cwebp for the browser case, and the non-WebP fallback stays a sane JPEG instead of ballooning into a PNG three times the size.

Uploading WebP is a trap that looks like an optimisation. The script defaults to jpg and prints a warning if you force webp.

This also satisfies the JPEG-over-WebP rule in [[open-graph-share-seo-geo]] — and its og:image:type invariant, since most unfurlers will receive image/jpeg, matching a declared image/jpeg.

Naming: BEM, Train-Case, ISO stamp last

Block__Element--Modifier_YYYYMMDDTHHMMSSZ.jpg

Plane__Work-Items--Empty_20260815T052358Z.jpg
Railway__Usage-By-Project_20260815T045242Z.jpg
Postiz__Memory-Breakdown--TWF_20260815T045242Z.jpg

This deliberately mirrors the house convention already in use for share imagery (ogimage__Lossless-At--Banner.jpg, per [[generate-consistent-og-images]]) — same BEM shape, same Train-Case segments. Do not fragment it.

Separators, each meaning one thing:

Block is the subject, not the asset type. Plane, Railway, Postiz — not Screenshot. Everything in the folder is a screenshot, so spending the most significant position on a constant wastes it.

The case rule (one principle, both conventions)

Lowercase for generic type tokens. Train-Case for proper nouns and named things.

That single rule explains both image conventions in this codebase, so neither is an exception to be “fixed”:

ogimage__Lossless-At--Banner.jpg
└─ type   └─ proper noun └─ named variant
   lower     Train-Case     Train-Case

Plane__Work-Items--Empty_20260815T052358Z.jpg
└─ product └─ named view └─ state
   Train-Case  Train-Case   Train-Case

ogimage is lowercase because it is a kind of asset, not a name. Screenshots carry no type token at all — the subject occupies the block, and the folder already says what they are.

Train-Case for the rest because these are proper nouns and named conventions, and reading Work-Items beats parsing work-items at a glance.

It is Title-Case, not capitalize-every-word. Minor words — articles, short prepositions, conjunctions — stay lowercase unless they lead or close a segment. The capitalised words are the ones carrying meaning, which is the whole point of the convention:

Image-of-Amy              not  Image-Of-Amy
Railway__Usage-by-Project not  Railway__Usage-By-Project
The-State-of-the-Art      not  The-State-Of-The-Art
A-Note-to-Self            not  A-Note-To-Self

Minor set: a an and as at but by for from in nor of on or per so the to v vs via with yet. First and last word in a segment are always capitalised, even when minor — A-Note-to-Self keeps its leading A.

The script applies this automatically, so pass --name however reads naturally and let it normalise.

On case-insensitive filesystems: Plane.jpg and plane.jpg collide on APFS but are distinct on the CDN — a real hazard in general, but not here. The script generates the markdown reference from the same string it writes the file with, so the case can never drift between them.

Timestamp: ISO 8601 basic format, UTC.

20260815T052358Z    ← basic format. This IS the standard, not a homemade variant.
2026-08-15T05:23:58Z  ← extended. Colons break URLs and Windows paths.
20260814-235021       ← local time, no offset. Ambiguous forever after.

Basic format exists precisely to solve separators-that-break-filenames. Use T and Z; don’t invent a lookalike. UTC because a local stamp with no offset can’t be resolved six months later and sorts wrongly against captures from another zone.

Semantics first, timestamp last is a deliberate trade. Filenames are a weak ranking signal, so the meaningful part leads. The timestamp still buys guaranteed uniqueness — which is what makes re-running safe, since two runs can never collide on a live URL.

Staging: per-repo and gitignored

Default is <cwd>/.image-staging/. Per-repo rather than global, so the processed image sits next to the content that references it and the provenance is obvious.

Add it to the repo’s .gitignore. The CDN is the source of truth for the served bytes; the staging copy exists so a re-run doesn’t re-download and so you can eyeball what was uploaded. Committing them doubles the repo weight for no benefit — the same reasoning that keeps .ideogram-candidates/ PNGs out of git.

.image-staging/

Emit modes

On our own Astro Knots sites, prefer lfm. Those sites do not use Astro’s markdown pipeline at all — LFM parses each entry and src/components/markdown renders it, so the ::image{…} directive is the richest target available (captions, sourcing, float layout). Raw HTML also survives, via a set:html fragment, so html works if you need dimensions.

Elsewhere, prefer html. Cumulative Layout Shift is a real ranking factor, and plain ![](…) gives the browser no way to reserve space — the page jumps as each image loads. Use md only where the renderer strips HTML.

Known gap, tracked not fixed: neither LFM path emits intrinsic width/height, so CLS is unsolved on our sites regardless of which you pick. Latent while no changelog has images. See [[Images-in-LFM-Render-Without-Intrinsic-Dimensions]].

SEO checklist, beyond alt text

• Dimensions on every <img> — reserves space, kills CLS
• loading="lazy" below the fold; omit it for the first image (lazy-loading the LCP element makes the metric worse)
• Responsive variants via ImageKit transformations rather than re-uploads: ?tr=w-800 — one upload serves every breakpoint
• Caption ≠ alt. A <figcaption> is read in addition to alt text, not instead. Don’t duplicate the string verbatim across both.
• Don’t use title as a substitute for alt — it isn’t announced reliably and doesn’t show on touch devices

Folder operations

ImageKit folders are created implicitly by the folder parameter on upload — there is no need to pre-create them. What you do need is a way to see and clear what is there:

node scripts/prep-images.mjs --list-folder  /self-host-stack/plane-first-deploy
node scripts/prep-images.mjs --purge-folder /self-host-stack/plane-first-deploy

--purge-folder is irreversible and deletes every file at that path. Use it for clearing test uploads and abandoned drafts, not for anything published.

The folder path is derived, never passed per-call: /<repo>/<slug>/. That mirrors where the content lives, so the CDN stays navigable and a whole entry’s imagery can be listed or cleared in one command.

Gotchas

1. sips cannot write WebP. It handles resize and JPEG fine; WebP needs cwebp (brew install webp). Since JPEG is the recommendation, this rarely matters.
2. Screenshots carry metadata. Window titles, device identifiers, sometimes location. The script runs exiftool -all= before upload. Don’t skip it on anything client-facing.
3. Re-running is safe, because of the timestamp. useUniqueFileName=false + overwriteFile=true would normally mean a re-run clobbers a live URL — that was true of an earlier <slug>-<NN>- scheme and was its worst property. The ISO stamp removes it: two runs produce two filenames, so nothing published can be overwritten. The cost is that repeated runs accumulate near-duplicates in the CDN folder; prune deliberately rather than relying on overwrite.
4. Retina screenshots are 2×. A “900px” macOS screenshot is often 1800px of pixels. The 1600px default is a reasonable ceiling for in-page content; raise it with --width when detail actually matters.
5. The ImageKit folder is derived, not configurable per-call — /<repo>/<slug>/. That is intentional: it mirrors where the content lives, so the CDN stays navigable.

Related

• [[open-graph-share-seo-geo]] — where og:image should be hosted, the JPEG-over-WebP rule this skill’s measurement confirms, and the og:image:type-must-match-bytes invariant
• [[generate-consistent-og-images]] — for generating share imagery; this skill is for screenshots of real things
• [[changelog-conventions]] — the surface these images most often land on
• [[maintain-splash-pages]] — the other main consumer
• content-farm/plugin-modules/image-gin — the Obsidian-side counterpart; src/services/imagekitService.ts is where the auth pattern comes from