Creative polish
Skill DKeken/codex-skills-alternative/skills/creative-polish
17 vendor-neutral Agent Skills (SKILL.md) reimplementing the Codex Creative Production + Product Design plugin workflows. Works with Claude Code, opencode, Cursor, Codex CLI, Gemini CLI, Copilot.
npx -y skills add DKeken/codex-skills-alternative --skill creative-polishAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 13 stars13 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Take a chosen visual to finished, production-ready state. Runs cleanup + refinement passes (fix artifacts, balance composition, color-grade, sharpen text, conform to spec/aspect) using image edit/regenerate. Use when the user says "polish this", "finalize", "clean up the image", "production-ready", "fix the artifacts", or has a selected visual that needs finishing. Final stage of the Creative Production set.
SKILL.md
4.1 KB, as published. Nobody here has run it
Creative — Generative Polish
Turn a selected draft into a final, production-ready asset. This is the LAST stage — it only runs on an already-chosen visual. Polishing before a concept is locked wastes effort; this skill assumes the creative direction is settled and now it's about execution quality.
When to use
- A concept has been chosen (from explore/scene/shot/ads/offer) and needs finishing.
- An image has fixable flaws: artifacts, bad hands/text, off composition, wrong aspect.
- Conforming a winner to final delivery specs (size, ratio, color).
When NOT to use
- The concept isn't locked yet — polishing many candidates is wasteful; lock first.
- You need new concepts → the relevant explorer skill.
- The image is fundamentally wrong (bad composition at the core) — regenerate upstream instead of patching.
Inputs (ask only what's missing)
- The selected source image (path or reference).
- Target spec: final aspect/size, where it'll be used, brand constraints.
- What to improve — specific issues, or run a full pass.
Workflow
- Diagnose the source. List concrete issues: artifacts, extra fingers/garbled text, composition imbalance, color cast, soft focus, wrong aspect, distracting background.
- Decide per issue: edit (inpaint/region fix via an image-edit tool) when only a region is wrong — it keeps the rest stable; regenerate (re-prompt with stronger constraints + source as reference) when the issue is structural.
- Apply passes in order: structural fixes → composition/crop → color grade → text legibility → final sharpen/export. Save iterations to
./creative/polish/<NN>.pngso you can roll back. - Conform to final aspect/size for the target medium.
- Show before/after and confirm it meets spec.
Worked example
Source: chosen hero from creative-shot, but has a duplicated keycap, blurry legend text, slightly cool cast, and is 1:1 but the hero slot needs 16:9.
issues: [duplicated keycap top-row] [legend text soft] [too cool ~-300K] [need 1792x1024]
passes:
1. edit (inpaint): remove duplicate keycap, rebuild row — region fix, rest untouched
2. regen-region: sharpen legend text, pass exact legends verbatim
3. grade: warm +250K, lift shadows slightly
4. crop/extend to 1792x1024 (outpaint left/right of product)
before: ./creative/shot/hero-34.png after: ./creative/polish/hero-final.png
verify: keycaps correct, text crisp, neutral-warm, 1792x1024 ✓
Quality bar
- No visible generation artifacts (hands, text, duplicated/merged elements).
- Exact visible text/wording preserved — never mutated during edits.
- Color and contrast are intentional and consistent with the brand/set.
- Delivered at the correct final aspect and resolution for its use.
Common pitfalls
- Text mutation — regenerating a region silently changes wording. Always pass text verbatim and re-check.
- Full-regen when an edit would do — re-rolling the whole image loses everything that was already right. Prefer region edits.
- Polishing unlocked concepts — finishing 5 candidates nobody chose. Lock first.
- No rollback — overwriting the source. Save numbered iterations.
Handoff
This is the terminal stage. Output the final asset(s) at spec. If the user wants the same treatment across a set, apply the same pass recipe to each and verify consistency.
Tooling
Needs an image tool with edit/inpaint + generate support (ideally outpaint for aspect changes). If only generation (no edit) is available, prefer targeted regeneration with the source as reference; if no image tool at all, output the diagnosed issue list + the exact fix prompts.