image-studio
Turn a business website link into a ready-to-use set of marketing images that show what the business actually sells. Reads the site, captures real proof of the work (client sites, product UI, demo elements) and frames it in browser and phone mockups, harvests its best existing photos, fixes the near-misses with Codex edits, generates whatever is missing with Codex (free, ChatGPT quota), rejects anything that looks like AI slop or could be any business, and crops each image to the size it will be used at. Delivers exactly N images into a new folder in Downloads. Use for "/image-studio", "images for this site", "marketing images for <link>", "give me N images for this business", "social images from my website", "find the best images on this site", "make images for this brand", "edit these images for Instagram", or any request for a set of brand or marketing images (not video: that is creator-studio).
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/adarsh350--image-studio)HTML
<a href="/listings/adarsh350--image-studio"><img src="/api/listings/adarsh350--image-studio/badge.svg" alt="findsafeskills safety rating" /></a>Install
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 241 lines
- ---
- name: image-studio
- description: Turn a business website link into a ready-to-use set of marketing images that show what the business actually sells. Reads the site, captures real proof of the work (client sites, product UI, demo elements) and frames it in browser and phone mockups, harvests its best existing photos, fixes the near-misses with Codex edits, generates whatever is missing with Codex (free, ChatGPT quota), rejects anything that looks like AI slop or could be any business, and crops each image to the size it will be used at. Delivers exactly N images into a new folder in Downloads. Use for "/image-studio", "images for this site", "marketing images for <link>", "give me N images for this business", "social images from my website", "find the best images on this site", "make images for this brand", "edit these images for Instagram", or any request for a set of brand or marketing images (not video: that is creator-studio).
- argument-hint: "<url> [count] [--use social|web|ads|mixed] [--proof <folder>] [--existing-only]"
- ---
- # Image Studio
- Link in, N finished images out. `<skill-dir>` is this skill's base directory. First time on a
- machine: `npm install` in `<skill-dir>` (Playwright; the Chromium it needs is shared with
- creator-studio, else `npx playwright install chromium`). Needs Node 20+, Python 3.10+ with Pillow,
- and the Codex CLI or the OpenAI VS Code extension signed in to ChatGPT.
- Read these before writing the brief. The quality of the set is decided there.
- - `<skill-dir>/references/playbook.md`: what to make. Placements and their specs, worked prompts
- for 12 business types, and the default mix for "N images".
- - `<skill-dir>/references/craft.md`: how to make and judge each image. The brief template, the
- prompt recipe, style anchors and the three-pass reject gate.
- - `<skill-dir>/references/industries/INDEX.md`: maps any business to one of 52 industry sections
- (benchmark brands, house style, slop tells, shot concepts, real-only shots), with a fallback for
- a business no section fits.
- **The bar:** every delivered image can go straight into a paid ad or onto a client's Instagram
- today. If it would look cheap next to the photos in `references/anchors/`, it isn't delivered.
- And the set sells: at least 60% of it shows the business's real work or product, so a stranger
- scrolling past could say what the business does (craft.md section 2).
- ## 1. Intake
- Need three things: the **link**, the **count**, and the **use**. If the invocation gave all three,
- don't ask. Otherwise ask once with AskUserQuestion, only for what is missing:
- | Question | Options |
- |---|---|
- | How many images? | 6 · 10 · 20 (Other for any number) |
- | What are they for? | Mixed social + web (Recommended) · Social posts · Website · Ads |
- `--proof <folder>` (or screenshots the user attaches or points to): real screenshots of the work,
- such as a product dashboard, a client site, a report. Read every file in it. Rank them by how
- clearly each shows one offer, how sharp and large it is, and whether it holds personal data
- (skip those unless the owner cleared them). Then use the best one or two per offer as `proof`
- slots (craft.md section 2). Copy any proof file the set uses into `$W/proof/`. Never generate a
- stand-in for proof the user said they will supply: leave the slot open and say so.
- Defaults when the user says "just do it": 10 images, mixed. Whose brand: a client site gets no
- Iyara context; Iyara's own site loads `iyara-labs` for the brief. If the count is over 20, say it
- draws on the ChatGPT allowance (about 35s and one message per generated image, plus retries) and
- continue.
- **Sizes by use** (each slot gets one):
- | Use | Size | Ratio |
- |---|---|---|
- | Feed post (Instagram, LinkedIn, Facebook) | 1080x1350 | 4:5 |
- | Square post / profile tile | 1080x1080 | 1:1 |
- | Story, Reel cover, TikTok | 1080x1920 | 9:16 |
- | Website hero, blog header | 1920x1080 | 16:9 |
- | Link preview (OG), ad landscape | 1200x630 | 1.91:1 |
- Mixed set: use the default set in playbook.md section 4. It includes the Google Business Profile
- slots, which take **real photos only** (Google's policy bars stock imagery, and a generated
- photo of premises misrepresents them). If the site has no real photo for a GBP slot, the slot
- becomes a feed post, and the reply lists the GBP shots the owner should take.
- ## 2. Read the site
- ```bash
- W="$LOCALAPPDATA/image-studio/runs/<slug>-$(date +%Y-%m-%d-%H%M%S)" # ~/.local/share/image-studio/runs on macOS/Linux
- node <skill-dir>/scripts/harvest.mjs "<url>" "$W/site" --pages 6
- ```
- Writes `site.json` (pages, headings, nav, meta, fonts, palette, every image with size and role),
- `home.png` and `images/`. Then:
- - Read `home.png` and the 6-10 largest `photo` images (contact sheet first:
- `python -I <skill-dir>/scripts/finish.py sheet "$W/site/sheet.jpg" <files>`).
- - Write `$W/brief.md` per craft.md section 1, including **Offers and proof**.
- - Sites behind bot protection or empty shells: say so, retry once with `--pages 1`, else ask for
- screenshots or a folder of images and continue from the brief.
- **Capture the proof.** For each offer in the brief, screenshot the real thing:
- the client sites it links to (`site.json` `pages[].outbound`, plus the work or case-study pages),
- and the UI or demo elements on its own pages (`pages[].ui`: a selector and a hint). View
- `home.png` first and pick only what is plainly the work, not navigation or a footer.
- ```bash
- node <skill-dir>/scripts/capture.mjs --batch "$W/capture.json"
- # capture.json: [{"url": "https://client-site.com", "out": "<W>/cap/client-desktop.png"},
- # {"url": "https://client-site.com", "out": "<W>/cap/client-mobile.png", "device": "mobile"},
- # {"url": "https://client-site.com", "out": "<W>/cap/client-full.png", "full": true},
- # {"url": "<url>", "out": "<W>/cap/demo.png", "selector": "<pages[].ui selector>"}]
- ```
- desktop is 1440x900 at 2x and mobile 390x844 at 3x; cookie banners and chat widgets are hidden.
- Read every capture: a half-loaded page, a surviving banner or placeholder copy is recaptured
- (raise `--wait`) or dropped.
- Work shown only as lazy-loaded demo videos (a case-studies page with no outbound links) is still
- proof: scroll the page so the `<video>` sources load, download them, and pull frames with
- `ffmpeg -ss <t> -i v.mp4 -frames:v 1 f.png`. Screen recordings almost always carry the mouse
- pointer. Pick the frame where it sits over flat colour and paint it out with that colour
- (cleanup, not an edit). If it covers copy or numbers, use another frame.
- ## 3. Shot list
- Find the business in `references/industries/INDEX.md` and read its section (INDEX.md says how to
- use each part and what to do when nothing fits). Its House style goes into the brief and every
- prompt; its Never list joins the reject gate; its Real only shots are never `generate`.
- Write `$W/shots.md` per craft.md section 2: N rows of `#, use, size, shows, concept, source, ref`.
- Assign showcase slots first (at least 60% of N, every offer once when N allows), from the proof,
- the section's Showcase concepts and playbook.md section 3. Then fill mood slots (at most 40%) from
- the Shot concepts.
- Post it in chat as a compact table and **carry on without waiting** (the user asked for a link-in,
- files-out flow). `--existing-only` means no `generate` rows: if the site lacks N good photos,
- deliver fewer and say how many it had.
- Harvested images are good for `existing` only when all of these hold:
- - they are photos, not graphics;
- - they are sharp;
- - they have no baked-in text;
- - they are big enough that the final crop upscales 1.5x or less.
- Mine big site photos before generating anything. One photo with a long side of 3000px or more
- often holds two or three deliverables at its natural subject points: the wide scene, a person, a
- detail. Each crop is its own `existing` row with its own `--focus`. Logos are never a slot.
- ## 4. Make
- - **Proof, capture, existing:** straight to finishing (or into a mockup).
- - **Mockup:** frame captures on the brand ground at the slot's delivery size, all in one batch:
- ```bash
- node <skill-dir>/scripts/mockup.mjs --batch "$W/mockup.json"
- # mockup.json: [{"layout": "pair", "out": "<W>/raw/02.png", "size": "1080x1350",
- # "imgs": ["<W>/cap/client-desktop.png", "<W>/cap/client-mobile.png"],
- # "bg": "<brand ground hex>", "url": "client-site.com"},
- # {"layout": "split", "out": "<W>/raw/05.png", "size": "1920x1080",
- # "imgs": ["<W>/cap/before.png", "<W>/cap/after.png"], "labels": ["Before", "After"]}]
- ```
- Layouts: `browser` (1 desktop capture), `phone` (1 mobile), `pair` (desktop + mobile), `split`
- (before + after), `element` (1 UI element). Options: `bg`, `fg` (label colour), `ground` (a
- texture image), `chrome` light|dark, `url` (the real domain, the only text drawn besides the
- labels), `margin`. Output is 2x the size, for `finish.py fit` to bring down. A `WARN` (capture
- upscaled past 1.5x) means a bigger capture or a different layout.
- - **Screen:** generate the device scene with a flat green screen (craft.md "Screen scenes"), then
- `python -I <skill-dir>/scripts/finish.py screen <scene> <capture> "$W/raw/NN.png"`.
- - **Edit and generate:** write prompts per craft.md section 3 into `$W/jobs.json` and run all of
- them in one batch:
- ```bash
- python -I <skill-dir>/scripts/codex_image.py --batch "$W/jobs.json" --workers 3
- # jobs.json: [{"prompt": "...", "out": "<W>/raw/04a.png", "aspect": "4:5",
- # "style_refs": ["<skill-dir>/references/anchors/hands-craft-fresh-pasta.jpg"]},
- # {"prompt": "extend the scene left and right...", "out": "<W>/raw/07.png", "aspect": "16:9",
- # "refs": ["<W>/site/images/007.jpg"]}]
- ```
- The script lints every prompt first and spends no quota if any prompt fails. It then wraps each
- brief in Codex instructions: a labelled spec, the size for the aspect, the highest
- quality, and the fixed realism and exclusion clauses. Anchors (`style_refs`) come from
- `references/anchors/manifest.json`, by the shot's look first and the industry second (craft.md,
- "Style anchors").
- Make two candidates per `generate` slot so the gate has a choice: the same concept, one job with
- `style_refs` and one without. Anchors pull the grade toward the site; the plain run often keeps
- more fine detail. Tested 2026-11-19: each won one slot. If both candidates fail, rewrite the prompt. Ask Codex for the slot's own aspect: the built-in tool returns about 1.5MP at that aspect
- whatever size is requested (4:5 comes back 1122x1402), which is enough for every slot up to a
- 1920x1080 cover. The result's `size` field shows what came back, and `qa.py` catches anything too
- small.
- No Codex (`codex_image.py --check` fails): don't use a paid image API. Fall back to
- `--existing-only` and say so.
- ## 5. Gate
- Run craft.md section 4 on every candidate:
- 1. `qa.py` (machine pass);
- 2. the contact sheet;
- 3. full size, with `finish.py zoom` on every hand, face and possible lettering.
- Retry rejects as the gate says: rewrite once, try a new concept once, then fall back to the
- next-best harvested photo. After finishing, run `qa.py` once more on `final/` at each delivery
- size (it takes file paths, not a folder: `qa.py "$W"/final/0[1-5]-*.jpg --size 1080x1350`), and
- look at the final contact sheet as one set.
- ## 6. Finish and deliver
- ```bash
- python -I <skill-dir>/scripts/finish.py fit <source> "$W/final/01-feed-<concept>.jpg" --size 1080x1350 --focus 0.5,0.4
- ```
- Pick `--focus` from what you saw (the subject's position, as fractions of width and height). The
- script warns above a 1.5x upscale: then pick another source or regenerate, don't ship it soft.
- - **Deliver:** copy `final/` to `~/Downloads/<slug>-images-<YYYY-MM-DD>/` (add `-2`, `-3` if it
- exists). Exactly N images, named `NN-<use>-<concept>.jpg`. Nothing else in that folder.
- - **Chat reply:** the folder path; a table of `file, size, shows, source (site page URL / capture
- of <url> / supplied proof / Codex edit / Codex generated)`; any offer that had no proof, as a
- request for screenshots; anything the user should check (an edited real photo, a slot that fell back);
- the real-only and GBP shots the site lacked, as a short list for the owner to take; and the work
- dir for re-rolls.
- - **Rights:** images harvested from a site may be licensed stock that is cleared for that website
- only. Say so in one line when the site isn't the user's own or the set includes harvested photos.
- - **Re-rolls** ("redo 04", "warmer", "two more"): reuse the latest work dir for that slug, change
- only those slots, overwrite them in the same Downloads folder.
- - **Prune** at the start of each run: `find "$LOCALAPPDATA/image-studio/runs" -mindepth 1 -maxdepth 1 -type d -mtime +30 -exec rm -rf {} +`
- ## Hard rules
- - Never generate the client's logo, product as sold, team, premises, certificates, screenshots,
- product UI, dashboards, client work, AI answers or metrics. Those come from the site (captured),
- or from the user.
- - At least 60% of the set is showcase, and every showcase slot passes the stranger test.
- - Google Business Profile slots are real photos only.
- - Never put text, lettering or logos in a generated image. Real pixels (captures, supplied
- screenshots, the site's own copy) may carry their own text.
- - Never alter what a capture or supplied screenshot shows: crop, frame and composite it, never
- edit its numbers or copy.
- - Never ship an image that failed the gate to make up the count. Deliver fewer and say why.
- - Paid tools need a flag first: the Higgsfield MCP, any image API other than Codex, any paid
- upscaler.
- ## Related skills
- - `creator-studio`: WHEN the ask is a video. It can take this skill's finished images as stills.
- - `social` / `instagram-content` / `linkedin-content`: WHEN the images need captions and a
- posting plan. They consume the Downloads folder this skill writes.
- - `brand`: WHEN the site has no usable visual identity and one has to be defined first.
- ## Self-tests
- `node scripts/harvest.mjs --self-test` · `node scripts/capture.mjs --self-test` ·
- `node scripts/mockup.mjs --self-test` · `python -I scripts/finish.py --self-test` (fit, sheet,
- zoom and screen) ·
- `python -I scripts/qa.py --self-test` · `python -I scripts/codex_image.py --check` (lint and prompt tests
- plus the Codex version; no quota used). All at once: `npm run check`.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.