make-product-photos

The engine composes, generates, judges and retries. This playbook is the e-commerce creative director around it: it decides which pictures the page needs, curates which real photos the engine sees, turns each shot into a precise prompt, holds every scene to what the product physically is, checks each output against the real product, and keeps the catalogue looking like one shoot.

What you get

• A shot list from the surface (store page, ads, social, email, launch page), each shot with its slot, composition, light, scene and ratio — agreed before anything is made.
• A house-style card derived from the brand's existing photos (when they can be seen), confirmed once and saved.
• The photo set as the page it fills: each photo with its use, the engine's status and score, this playbook's verdict (keep / maybe / reject + named reason), and who checked it (by eye, or engine only with a checklist for you).
• "What we could not make and why", and the credits actually charged.

Before you start

Hard rules:

1. Honesty: a scene may show only what the product physically is and does — its size, contents, finish, colour and how it is used — from the product record, the real photos and the human's confirmation. Never a result unless it is a signed-off claim (../../../skills/marketing/claims-gate/SKILL.md), never a feature it lacks (references/fit-and-honesty.md §2 is the one list).
2. References: only real photos of this exact product, size and colour; never text graphics, other sizes, duplicates or earlier generated images.
3. Money: nothing generates before the marketer's yes to the engine's quotes, shown one by one exactly as given — never your own sum (../../../skills/marketing/spend-gate/SKILL.md, branch A); the hero runs first; a re-shoot is quoted again.
4. People: any shot with a person or a body part (hands included) needs the human's own attestation; no real likeness; no children.
5. Keepers: only the human's picks are approved into the brand kit; nothing is uploaded to a store.
6. Honest review (../../../skills/marketing/honest-review/SKILL.md) and plain words: never name a playbook, tool, id, model or vendor to the marketer.

Setup:

0. Workspace: every file (brand/photo-style.md, the ads/batches/ request lines) uses the brand coworker scope (scope: {type: "agent", agent_id: <coworker_agent_id from the brand_read list>}); no coworker → the default workspace with every path prefixed brands/<brand-slug>/, and tell the human.
1. Connection. No gooseworks: tools (after one gooseworks:tools_load for a missing group: brand, media, files) → say only: "Connect GooseWorks first — I make your product photos through it."
2. gooseworks:get_started with wants: "image" and the marketer's words (there is no photos value); its answer settles setup and the brand only — a card → wait for the press, then Step 1 (this is product photography, not ads). gooseworks:account_whoami once: features.product_photos on, credits available.
3. Two modes, decided now:
   • Can see images (a shell host downloads the reference photos and outputs and views them): derive the style card, check outputs by eye.
   • Cannot see images (chat hosts get links, not pixels): no style card is derived — the marketer picks their best current hero as the anchor; outputs are compared field by field against the reference's automatic read (Step 7) and labelled "not checked by eye".
4. Mode: new product photos · new colourway · seasonal refresh (lifestyle only; the hero never changes) · imagery request from the ads side (Product imagery needed: for make-product-photos · … lines in ads/batches/*.md in the brand workspace; answer under each: Product imagery ready: <photo links> · <YYYY-MM-DD>) · packaging changed (retire and reshoot) · pre-launch with only a render (Step 1).

Steps

1. Fit check — can the studio photograph this honestly?

Reads: gooseworks:brand_read lean (sections: ["summary","kit","products","learnings"]; product_ids for the full record incl. description, notes, ai_creative_notes, listed_at) — a lean read returns every must and dont (photo "Don't …" rules, the claims sign-off); the saved house-style card (brand/photo-style.md, Step 3); kit.brandType; the product's type. Load references/fit-and-honesty.md §1.

• Map to apparel / beauty / cpg; hard goods → cpg with the risk said; electronics, furniture, jewellery, eyewear → lifestyle only or a real shoot; food service, hospitality, real estate → decline.
• Software, apps, services: one line — "this makes photos of physical products; screenshots and mockups are made as ad graphics instead" — and ask once whether there is a physical item (device, packaging, merch).
• Pre-launch, render only: with the human's yes, create the product if missing (gooseworks:brand_update products) and add the render as a product photo (Step 2, "getting a new real image in"; tag it render; it becomes the kit hero only if the human chooses); the render itself is the honest studio hero, unaltered; generate lifestyle only, at the render's angle, no screens or app UI; label the set "pre-production, render-based" and tell the human to caption launch-page images as renderings; ask for CAD exports for other angles; keep it out of the ad kit unless the human chooses.

2. Curate the real photos — the biggest lever

Run atoms/source-audit/ATOM.md. Reads: the product's photos (gooseworks:media_list scope: "product"; media_analyze read; your own eyes in "can see" mode). Writes: the reference inventory.

• Sort every photo: product-only, this size / colour (usable) · text graphic or claim overlay (never a reference) · other size, width or variant · duplicate (same content_revision — it takes the most suspicious class of any copy) · possibly generated (an earlier GooseWorks output — match against photos_read outputs of every status by media_asset_id / final_image_url and the full kit's generated image URLs — or a generator filename such as hf_…, pomelli-image, dalle, mj_, Gemini_Generated_Image, ChatGPT Image, firefly, leonardo). source.type: brand_catalog only means "came from the store" — stores hold AI images too (for example, a watch-band brand's newest colourways). When photos are unaudited, the product is newly listed, or no original filename is visible (re-hosted product-N-M copies), ask "were any of these made with an AI image tool?"
• Getting a new real image in (a missing angle, new packaging, a render, a colourway's photo), always on the human's yes: atoms/source-audit/ATOM.md §8. An upload becomes the product's hero in the brand kit only when the human says this photo should be the main image; new packaging → replace the product's photos and archive old approved outputs with no reason (Step 9).
• More than one size or format and no variants → ask "which size does this page sell?" first.
• Analysis blocked_budget or empty → "unaudited": ask the human what each photo shows.
• Rights: the catalogue photos are marked unverified — one line: "these are yours to use?"
• Dimensions: notes, ai_creative_notes, the label (OCR is a hint, not ground truth), or ask.
• Offer only the slots the usable photos support; for each missing one, name the unlock ("a phone photo of the back on a white sheet"; pre-launch: "export the back view from CAD") and a substitute (e.g. the hero in a second crop).

3. House style

Load references/shot-list-and-surfaces.md §3. A saved card (brand/photo-style.md in the brand workspace, or the brand document "Photo house style — <Brand> — <date>") → show it, ask "still right?", use it in both host types, and save the changed version if it changes. No saved card and "can see": derive the card from the brand's existing product photography and approved earlier outputs (gooseworks:photos_read status: "approved"), ask "match or deliberately upgrade?", confirm it once. "Cannot see": the marketer's chosen anchor hero plus the kit's aesthetic stand in; no card is saved. References vs anchor: only this product's best real product-only photos go in reference_image_urls. An approved generated hero is a style anchor only — describe its angle, light, crop and background in the prompt, never pass it as a reference (errors compound). Never a sibling product or colourway as either (colour bleed).

4. Propose the shot list — slots, not styles

Run atoms/shot-list/ATOM.md (references/shot-list-and-surfaces.md §1–2). Stack from the surface (for "the new colourway": read the products with detail: "full" in Step 1, sort by listed_at, confirm — on a near-tie, ask which); apply the brand's saved "Don't …" photo rules; honesty pass on every scene (Hard rule 1; skincare lifestyle shows the product, not a perfected face); host devices shown accurately (right model, neutral screen, no invented UI); colourways as variants or sibling products (find the new one by listed_at and confirm). Infographics and text over images are ad graphics, not this job. Ask only what the brain cannot settle; let the marketer edit.

5. Build the calls and quote — the spend gate

Per shot: one call — product, variant_id if any, category, count 1 or 2 (always set it; retries are shared per call), the shot row written into prompt (composition, light, background, copy space, what must not appear), up to six curated reference_image_urls (same-angle product-only first), controls only with keys the engine is known to honour, quality (offer medium for lifestyle). gooseworks:photos_generate dry_run: true for each; show each quote one by one exactly as given (no total of your own); people or hands → the human's attestation in plain words. Wait for the yes.

6. Hero first, then the rest

Submit the hero call alone. Read it back (gooseworks:photos_read generation_id): the ratio, quality and overridden_controls actually used, and run the fidelity check on the hero — a blocker, or controls the shot needed dropped, stops the run and goes to the human before the rest is spent. In "cannot see" mode, show the hero link with its three-item checklist and ask "carry on with the rest?" before submitting the other calls (same quotes, no new price). Poll gooseworks:job_get (kind: "photo_generation") for a bounded time; still queued → "I'll check when you're back" and end the turn — never resubmit (it pays again). workflow_error with nothing charged is an infrastructure failure: offer a fresh quote; quality failures carry final_defect_code / final_defect_message.

7. Check every output against the real product

Run atoms/fidelity-check/ATOM.md (follows ../../../skills/marketing/honest-review/SKILL.md; rubric in references/fidelity-rubric.md) on every passed and flagged output — final_image_url, quality_score, final_defect_*, media_asset_id. Honesty first, then parts, logo and text (by eye; OCR is a hint), silhouette, colour, scale, hands, physics, background, finish, house style. "Cannot see": gooseworks:media_analyze on the output's media_asset_id and on the same-angle reference, and compare them field by field — part and item counts, the readable label text, colour words, which product is shown — flagging each mismatch specifically ("reference label reads 'DIVE DEEP'; output reads 'DIVE DEEEP'"); label the photo "not checked by eye" and give a checklist aimed at what the reads could not settle. A re-shoot is quoted again, with the engine's suggested_control_change offered.

8. Present the set and the keeper pick

Lay it out as the page: slot, photo, use, engine status and score, verdict and reason, who checked it; recommend the page order (next to the current live hero when one exists). Say "generated from your product photos". Report the credits GooseWorks charged per call (credits_charged), as it reports them.

9. Record the decision

• Keepers the human picks: gooseworks:photos_update approve — on their words only (approved photos flow into ad generation).
• Rejects: archive with no reason — the engine turns any reason into a brand-wide "dont" the ad and video engines read, so the agent never writes one. A durable product rule the human states in their own words ("Don't show Dive Deep with a pump top — it has a flip cap") goes through ../../../skills/marketing/save-learning/SKILL.md (text starting [photos YYYY-MM-DD], one idea each, own superseded ones retired). Fine-but-unpicked photos are left alone.
• Old approved photos the human retires (packaging changed, product changed): archive with no reason.
• The confirmed style card: brand/photo-style.md saved as the brand document "Photo house style — <Brand> — <YYYY-MM-DD>" through ../../../skills/marketing/save-to-brain/SKILL.md (versioned: supersede). A rule the marketer states in their own words ("never a white background for this range") → ../../../skills/marketing/save-learning/SKILL.md, text starting [photos YYYY-MM-DD].

Approval gates

1. Shot list — agreed before any quote.
2. Spend — the yes to the engine's quotes; the hero first; every re-shoot quoted again.
3. Attestation — for any person or body part, the human's own confirmation each time.
4. Keepers — the human's pick before anything is approved into the brand kit. Putting photos live is the human's act in their own store admin.

Output and write-back

The set in GooseWorks (approved keepers linked to the product, available to the ad and video engines); the presented page in chat; the style card as a brand document and any rule the human states, through save-learning, on the human's word.

Done when

You would put the keepers on your store page today, unedited — and no photo this playbook marked "keep" fails your own fidelity look (zero false passes): right size, shape, parts, logo, text, colour, finish and scale, nothing implied that the product does not do.

Out of scope

• Ad creative, infographics, dimension callouts, text over imagery, screenshots and device mockups — make-static-ads.
• Product video — make-video-ads. Social posts — run-organic-social.
• Uploading to a store, marketplace or feed — the human's own act. Marketplace listing image sets — optimize-listings (backlog).
• Retouching or background-removing real photos; colourway line-ups in one frame.