Templatize Page

Templatize Page

Given a brand's lookbook / collection / editorial URL, build an Astria pack whose prompts let the user re-shoot that page's looks with their own workspace references.

There are two variants:

• shoe-swap — barefoot-edit each lifestyle photo (remove the footwear) and build a prompt that swaps in a shoe reference. Footwear-only.
• silhouette — silhouette-edit each photo (keep only the posed figure) and build a prompt where the user picks which items to swap with references (model, shoes, outfit, bottoms, background, accessories). Whatever they don't swap is described in the prompt text instead.

Both variants end with one new pack (titled after the page), N new pose tunes (one per surviving lifestyle image), and N new prompts inside the pack.

Two Python scripts ship with this skill in scripts/ (stdlib only). They call the bundled astria CLI for every API operation, so authentication and workspace scoping are handled for you — see the astria-api skill.

The ${CLAUDE_SKILL_DIR:-$HOME/.codex/skills/templatize-page} expansion in the commands below resolves the skill directory in every runtime: Claude Code exports CLAUDE_SKILL_DIR; the Codex-based Astria chat agent doesn't set it, but installs this skill at ~/.codex/skills/templatize-page.

Workflow

1. Choose the variant

Ask the user with your ask-user question tool (AskUserQuestion in Claude Code, ask_user in the Astria chat agent; single select) before scraping:

• Shoe-swap — replace the footwear in every shot.
• General pose (silhouette) — keep the pose, swap any items they choose.

2. Scrape the page

python3 "${CLAUDE_SKILL_DIR:-$HOME/.codex/skills/templatize-page}/scripts/scrape.py" <url>

Emits JSON: {"title": "...", "url": "<final url>", "images": [{"url": "...", "alt": "...", "width": ..., "height": ...}, ...]}. The page <title> (or og:title when set) is the pack title — trim marketing cruft (price, "✓️", site name) to the product / collection name. Capture both — title for the pack, images for the next step.

3. Classify each image (drop packshots)

The scrape returns every visible image — including product-only shots (a shoe on a plain background, no model) that we do NOT want as poses. Filter visually:

1. Download every candidate to /tmp/templatize/ so they can be read locally:

   mkdir -p /tmp/templatize && i=0
   for url in <url1> <url2> ...; do
     i=$((i+1)); curl -sSL -o "/tmp/templatize/img_${i}.jpg" "$url"
   done
   

2. View each /tmp/templatize/img_*.jpg image. For each one decide:
   • Keep when a person is visible wearing or holding product (a lifestyle / editorial / on-model shot).
   • Drop when the frame is a product alone — a single shoe / sandal / handbag on a plain or studio backdrop with no human figure. Also drop logos, banners, swatches.
   • When in doubt, keep. A borderline cropped-from-the-knees shot is fine; the edit step will normalize it.
3. Build the list of surviving original URLs (the public URLs — the scripts pass them to Nano Banana). Keep the local paths too — the silhouette variant reads them again to compose prompts.

4. Templatize — branch on the variant

4a. Shoe-swap variant

python3 "${CLAUDE_SKILL_DIR:-$HOME/.codex/skills/templatize-page}/scripts/templatize.py" \
  --pack-title "<page title>" \
  --pose-image-url <surviving_url_1> \
  --pose-image-url <surviving_url_2> \
  ...

The script: looks up the workspace's first shoe/sandal reference (astria tunes list --name shoes --name sandals, aborts if none); creates the pack; then per image barefoot-edits the photo, makes a pose faceid tune, and creates a <faceid:pose><faceid:shoe> swap prompt assigned to the pack.

4b. Silhouette variant

The silhouette edit strips the figure down to its pose, so the prompt has to rebuild everything else — references for the parts the user wants swappable, plain text for the parts they don't.

i. Ask which items to make swappable — one page-wide ask-user question with multiSelect: true. The references are shared across every pose, so this is asked once, not per image. Offer these categories:

• Model / person
• Shoes
• Top / outfit (shirt, dress, jacket)
• Bottoms (pants, tights, skirt, shorts)
• Background / setting
• Accessories (bag, jewelry, eyewear, headwear)

Checked → swapped with a reference tune. Unchecked → described in the prompt text.

ii. Pick a reference tune per checked category — for each checked category, list the workspace's tunes of that class and ask the user which one to use (ask-user question, single select; batch up to 4 categories per call):

If a checked category has no matching tune, tell the user and treat it as unchecked (describe it in text instead).

iii. Compose a prompt per image — look at each surviving image again and write a final prompt containing the literal placeholder {pose}:

Reproduce the same {pose} pose, <short pose description>, <references + descriptions>

• {pose} stays literal — templatize.py substitutes the new pose tune id.
• Add a brief pose description from the image (e.g. "leaning her hand on her knee").
• For each checked category: <faceid:{ref_tune_id}:1> {noun}.
• For each unchecked category visibly present in this image: a short text description (e.g. "green tank top", "white studio background"). Skip categories not present in the image.

Example (model + shoes checked; top, accessories, bottoms, background described):

Reproduce the same {pose} pose, leaning her hand on her knee, <faceid:4425929:1> woman with <faceid:4363887:1> shoes, green tank top, apple in-ear headset with wire, short blue tights, white studio background

iv. Write the spec — write /tmp/templatize/spec.json, a JSON array of {"url": "<original image url>", "prompt": "<composed prompt with {pose}>"} objects, one per surviving image.

v. Run the script

python3 "${CLAUDE_SKILL_DIR:-$HOME/.codex/skills/templatize-page}/scripts/templatize.py" \
  --mode silhouette \
  --pack-title "<page title>" \
  --spec /tmp/templatize/spec.json

The script: creates the pack; then per image silhouette-edits the photo, makes a pose faceid tune, substitutes the new pose tune id into the prompt's {pose} placeholder, and creates the prompt assigned to the pack.

Both variants accept --workspace <id> to target a specific workspace; otherwise the astria CLI's configured workspace is used. They also accept --aspect-ratio <ratio> to force one output ratio for the whole run — by default each source image is downloaded, measured, and the pack prompt rendered at the closest standard ratio. Each prints one status line per artifact to stderr and a final summary block to stdout with the new pack ID and per-pose prompt_id / pose_tune_id pairs.

5. Report

Surface the new pack to the user — quote the pack title + ID, the number of prompts created, and point them to /packs/<pack_id>.

Worked examples

Shoe-swap. /templatize-page https://www.gentlesouls.com/lookbook/spring-2026

1. Variant question → shoe-swap.
2. scrape.py → title="Gentle Souls Spring 2026 Lookbook", ~14 images.
3. Download all 14, view each — 5 packshots dropped, 9 on-model kept.
4. templatize.py --pack-title "Gentle Souls Spring 2026 Lookbook" --pose-image-url ... (9 URLs).
5. Report: "Created pack Gentle Souls Spring 2026 Lookbook (id 4821) with 9 shoe-swap prompts using reference Brown leather sandal (id 4457109) — /packs/4821."

Silhouette. /templatize-page https://brand.com/editorial/resort

1. Variant question → general pose (silhouette).
2. scrape.py + classify → 6 lifestyle images kept.
3. Checkbox → user checks Model and Shoes.
4. Reference questions → Model = tune 4425929 "Studio model — Maya", Shoes = tune 4363887 "Tan block-heel sandal".
5. Per image, view it and compose a {pose} prompt — refs for model + shoes, text for the unchecked top / bottoms / background. Write spec.json.
6. templatize.py --mode silhouette --pack-title "Resort Editorial" --spec /tmp/templatize/spec.json.
7. Report: "Created pack Resort Editorial (id 4830) with 6 general-pose prompts — /packs/4830."

Notes

• The barefoot / silhouette edit prompts are deliberately conservative — they keep the pose so the resulting tune captures the original styling minus the swapped-out parts.
• tune[name] is hardcoded to pose (every output should be usable as a pose ref regardless of the original subject).
• The pack uses model_type=faceid because the prompts inside reference faceid tunes (pose + the swapped references). Generation defaults: resolution=2K, num_images=1. The barefoot / silhouette edit leaves aspect_ratio unset so it follows the input image (Gemini rejects aspect_ratio=auto); the pack prompt is rendered at the source image's aspect ratio — templatize.py downloads each source image, measures it, and snaps to the nearest standard ratio (1:1 2:3 3:4 4:5 9:16 5:4 4:3 3:2 16:9 21:9), defaulting to 3:4 when the image can't be read. Pass --aspect-ratio 3:4 to force one ratio for the whole run.
• Astria dedups prompts by (text, seed) within a tune, so an identical edit instruction on two source photos would collapse onto one prompt. templatize.py gives each edit a per-image seed (derived from the source URL) — never append a marker to the prompt text, Nano Banana may render it into the image.
• If the page is gated (Shopify login wall, paywall) scrape.py returns empty images — fall back to asking the user for direct image URLs.