AGENTS.md
Guidance for coding agents working on the Gumroad product image producer role.
Project Overview
The product image producer creates Gumroad product-page images from real product artifacts. The role should prove what buyers receive: PDF excerpts, template previews, file inventories, screenshots, cover variants, galleries, and section-specific visuals. Generated imagery is secondary and must remain product-specific.
Source Files
- agents/gumroad-product-image-producer.md defines the role.
- skills/gumroad-product-image-producer/SKILL.md defines artifact-first rules and insertion notes.
- products/
/product.json supplies product facts and claims to avoid. - products/
/build/manifest.json supplies rendered artifact paths. - products/
/assets/cover.png and assets/page-images/ store accepted images. - gumroad_result.json or product view output supplies live page media when editing a published product.
Setup Commands
bashUV_LINK_MODE=copy UV_PROJECT_ENVIRONMENT="$HOME/.cache/gumroad-codex-agent-venv" uv run python scripts/product_artifacts.py render products/<slug>/product.json
UV_LINK_MODE=copy UV_PROJECT_ENVIRONMENT="$HOME/.cache/gumroad-codex-agent-venv" uv run python scripts/product_artifacts.py qa products/<slug>
Rerun render and QA after adding or replacing cover and preview assets.
Change Rules For Agents
- Read product content before planning images.
- Prefer actual PDF pages, template excerpts, HTML screenshots, file manifests, and ZIP contents over generated scenes.
- Map every image to a description section instead of dumping a gallery near the top.
- Add short explanatory copy before inline images so the image proves the surrounding claim.
- Use image generation only when the owner asks or artifact-based proof cannot cover the need.
- When using image generation, make the prompt explicitly request the configured image model and reject generic AI art.
- Preserve existing product copy, discount banners, covers, files, and coupons unless the owner explicitly asks to change them.
- Verify live body images in a browser when the task updates a published page.
Visual Standard
Use high-resolution, readable, product-specific images. Avoid generic laptops, robots, glowing AI icons, abstract gradients, money imagery, fake dashboards, fake testimonials, fake logos, distorted hands, blurry output, and unreadable generated text.
Verification
Inspect at full size and thumbnail size. Confirm images prove real product files, do not overpromise, and do not contain private data. Run product QA to catch blank, undersized, or malformed image assets.
Security Notes
Do not expose buyer data, private account ids, private media URLs, local absolute paths, credentials, unreleased product URLs, or hidden operational notes in images or captions.
PR Checklist
- Product facts were read before image planning.
- Accepted images are artifact-first or clearly justified.
- Images are mapped to page sections.
- QA was rerun after image changes.
- Live page edits were explicit and verified.