# Output spec ## Deliverables | File | Format | Size | Notes | | --- | --- | --- | --- | | `covers/cover-01.png` ... `cover-06.png` | PNG | 1080 x 1920 (9:16) | One per title, the exact title and kicker drawn locally. Upload each as the custom cover of its Reel or TikTok | | `covers/manifest.json` | JSON | | Title boxes, title size, contrast and a fingerprint of each cover. Read by the check and the grid scripts | | `grid.png` | PNG | 3 tiles wide, 2 rows (1088 x 964 at the default 360 px tile) | The set as followers see it on the profile grid: 3:4 crops, newest first. `grid.png.json` beside it | | `preview.webp` | WebP | at most 2048 px wide | Optional, from `grid-preview.mjs --webp`, for a message or a web page; aim for 150 to 400 KB | | `art/` | PNG | 1088 x 1936 finals, 720 x 1280 drafts | The text-free art, kept so the titles can be redrawn for free | | `series.json` | JSON | | The plan: look, style, titles, subjects, stills, art paths | | `covers/guides/` | PNG | 1080 x 1920 | Optional, from `compose-covers.mjs --guides`: each cover with the zones drawn over it, for checking by eye | Default settings: `openai/gpt-image-2` for covers without references, `openai/gpt-image-2/edit` with references (a still, cover 1 as the look anchor, the approved draft; at most 3); drafts at quality low, 720 x 1280; finals at quality medium, 1088 x 1936; `output_format` png. Titles in Helvetica Neue Bold (DejaVu Sans Bold on Linux) unless you pass a font file. ## Approximate cost (GPT Image 2 on fal, token-billed; read 2026-11-09) | Step | 6 covers | 3 covers | 9 covers | | --- | --- | --- | --- | | Drafts at low (about $0.005 each, plus $0.01 per still) | about $0.03 | about $0.02 | about $0.05 | | Finals at medium after drafts (about $0.10 each, plus $0.01 per reference) | about $0.71 | about $0.35 | about $1.07 | | Finals with no drafts first | about $0.65 | about $0.32 | about $0.98 | | One re-roll of a final | about $0.12 | about $0.12 | about $0.12 | | Titles, grid, checks | free | free | free | A still from the creator's own video adds about $0.01 to each call for that cover. `build-requests.mjs --stage estimate` prints the total for the real plan; quote it and get one yes before the first paid call. fal returns no token count with the image, so check the fal dashboard after the run; fal's published portrait prices suggest the real charge is at or a little under the estimate. ## Checklist Technical (`scripts/check-covers.mjs`, exit code 0): 1. `series.json` has no `FILL:` left; 3, 6 or 9 covers numbered in order; every title 1 to 6 words and at most 32 characters. 2. Every cover is a 1080 x 1920 PNG and has not changed since it was composed (its fingerprint matches the manifest). 3. Every cover was composed on the art the plan names now, from the final art (not an upscaled draft). 4. Every title and kicker matches the plan letter for letter (in capitals when `style.case` is `upper`), in at most 3 lines. 5. Every title and kicker sits inside the safe title box: inside the 3:4 grid crop, out of the top 15 percent, the bottom 35 percent, the right 150 px and the left 90 px. A title outside the old 1:1 square crop is a warning only. 6. Titles are at least 96 px, and the whole set shares one size (unless drawn with `--per-cover-size`). 7. Contrast between the title and what is behind it is at least 3:1. Busy art behind a plain or shadow title is a warning. 8. A still marked `still_has_person` has consent recorded (`series.people` `real-with-consent`, `series.consent_confirmed` true). 9. With `--grid`: the grid was made from the current covers and its tiles are 3:4. With `--preview`: a WebP, at most 2048 px, ideally 150 to 400 KB. The check reads the title boxes from the manifest that `compose-covers.mjs` writes and fingerprints each cover against it. It cannot check a title drawn in another app: compose the titles with the script. By eye (open every cover at full size, then the grid): 10. **One series**: same look, light direction, camera angle and finish on all six; the subject at about the same size and place; the palette pattern as planned. 11. **No stray text in the art**: no letters, numbers, signs, labels, logos or watermarks, other than lettering on the creator's own product in a still. 12. **Clean art**: true-to-life scale, natural hands, no melted cutlery or doubled objects, no extra people. 13. **Every title read letter by letter** against the plan, on the cover and on the grid. 14. **Readable on the grid**: at grid size, every title can be read, and the covers are told apart at a glance. 15. **Subject inside the grid crop**: the hero thing is not cut off on the 3:4 tile. 16. **People**: none unless planned. A fictional person resembles no real or famous person and is an adult. A person from a still is exactly themselves (identity, age, skin tone, hair, features, body), with consent recorded, and the cover suggests no endorsement, result or reaction that did not happen. 17. **Honest**: the cover shows what the episode has; no invented results, prices, awards or claims. ## If a check fails Title, layout, colour, case, treatment and grid problems are free to fix: edit `series.json` and recompose. Art problems: at most 3 attempts per cover, change one thing per attempt and say which (`references/prompt-patterns.md`, "Failure modes"), compare each attempt with cover 1, then deliver the best and say plainly what still differs.