Your AI Assistant Edits Photos by Guesswork—Someone Wrote It a 24KB Darkroom Recipe

You ask Codex to edit a photo, and it either cranks the exposure into a fluorescent light or twists the saturation into neon colors—like tossing a palette into a washing machine.

The problem isn't that the model is dumb. The problem is: no one told it what the editing priorities are.

Fix light or adjust color first? Can you change a person's face shape? Can the sky be replaced? Should the composition be cropped? Every choice affects the final result, but the AI assistant defaults to pure guesswork.

LumaRescue is here to end this guessing game. It's not an editing tool; it's a darkroom recipe installed for Codex and Doubao Desktop—a 24KB SKILL.md that breaks editing into hard anchors, soft anchors, four-tier output, and master-level cropping, turning your AI assistant from "guessing" to "understanding."

Four-Tier Output: From "Clean" to "Stunning"

LumaRescue's core design is four-tier progressive output, not simply editing one photo then handing it over:

The first two tiers are "darkroom craftsmanship"—exposure curves, tone mapping, cropping composition, each step reproducible. The last two tiers are "AI imagination"—using generative capabilities to supplement information missing from the original photo.

Key rule: The beauty version must be noticeably more beautiful. If it's just cleaner but barely changed, LumaRescue's Beauty Failure Rule will deem it a failure and require a redo—add at least three stronger operations: bolder cropping, clearer main light direction, cleaner subject/background separation.

This is not "close enough" editing; it's "redo if not stunning enough."

Hard Anchors vs Soft Anchors: What Can Change, What Cannot

The biggest fear in editing isn't doing a bad job, it's overdoing it—face changed, text removed, building skewed. LumaRescue solves this with a anchor grading system:

Hard anchors (must preserve): Person identity, facial structure, expression, pose, clothing, important objects, readable text, logos, artwork content, artifact facts, building identity, landmarks, core scene layout.

Soft anchors (can change): Light quality, time-of-day feel, sky drama, atmospheric depth, background clutter, non-critical colors, cropping, aspect ratio, lens feel, bokeh, local contrast, texture.

The brilliance of this grading is: it's neither "keep everything" nor "let it all go." Many editing prompts either over-preserve (not even daring to touch the original composition) or over-release (even replacing faces). LumaRescue explicitly tells the AI: lock identity and facts, free light and atmosphere.

Example: A backlit portrait where the face is a silhouette. Traditional editing might brighten the face to match the background brightness—information preserved, but light layers lost. LumaRescue's logic: first decide the main light direction, then decide which shadows should retain depth. The face can be brightened, but not evenly—simulate a side fill light effect, preserving light-dark transitions.

Master-Level Cropping: Your Phone's Composition Is Not a Commandment

Most people's photo composition is "raise the phone and press the shutter"—horizon tilted, half a meter of empty space above the head, half a person cut off at the edge. LumaRescue's Master Composition Rule says it plainly:

Don't assume the frame is correct just because the camera captured this frame.

Every image defaults to master-level cropping evaluation—compare at least four composition options: improved original aspect ratio, tighter subject crop, expressive aspect ratio (panorama/vertical/square/cinema widescreen), negative space composition. Only keep the original composition when it's already the strongest choice.

This is completely opposite to most editing tools. Traditional tools preserve the original composition by default, cropping is optional. LumaRescue defaults to recomposing; preserving the original composition requires justification.

7 Style Directions + Master Anchors

photographic-style-playbook.md defines 7 style directions, each with corresponding master anchors—not to imitate master styles, but to translate the master's technical decisions into universal operations:

Each style direction requires choosing composition first, then color. Only changing color without changing composition is considered a half-finished product by LumaRescue.

Competitive Landscape: Editing Track in the Skill Ecosystem

LumaRescue is not the only Agent editing skill, but its approach is the most distinct:

claude-image solves the issue of "Agent doesn't know how to use GPT Image 2"—Agent trained on old models still writes prompts like "4K, ultra detailed, masterpiece" which are ineffective for GPT Image 2. LumaRescue solves the issue of "Agent doesn't know how to edit photos"—not an API calling problem, but an aesthetic decision problem.

The two are complementary: claude-image teaches how to correctly call image generation, LumaRescue teaches how to judge the result after generation.

Cost Awareness

No License. The repository doesn't declare any open-source license, meaning legally you only have the right to view, not use, modify, or distribute. For an early project with 34 stars, this might be because the author hasn't added one yet, but it also means the project could become closed-source or change licensing at any time.

Zero Code, Pure Prompt. The entire repository has no executable code—SKILL.md is a prompt, references are a knowledge base, agents/openai.yaml is a registration file. Its capability depends entirely on the host Agent (Codex/Doubao)'s image processing ability. If the host lacks image editing tools, LumaRescue can only output textual descriptions of "what should be done" rather than actual edited images.

Sole Contributor, Bus Factor = 1. All 4 commits come from justlovemaki, no other contributors. Project sustainability depends entirely on one person.

The Double-Edged Sword of Beauty Failure Rule. "Redo if not beautiful enough" sounds good but in practice means more token consumption and longer waiting times. Each redo is a full image generation call, potentially doubling or tripling costs.

Gray Area in Style Anchor Copyright. Although SKILL.md explicitly says "don't copy the exact style of living photographers," using names like Ansel Adams, Avedon as anchors might still lead the AI model to lean toward imitating their signature styles. In commercial usage scenarios, this poses potential risks.

Judgment

LumaRescue's true value is not in what photos it edits, but in defining a new Agent Skill paradigm: not giving tools, but giving standards.

Traditional Skills teach the Agent "how to do" (which API to call, what parameters to pass). LumaRescue teaches the Agent "how to judge"—what is good light, which composition is stronger than the original, what to keep and what to let go. This is an upgrade from "operation manual" to "aesthetic standard."

But 34 stars, no license, pure prompt, sole contributor—it's more of a design document than a mature tool. If you use Codex or Doubao for photo editing tasks, it's worth cloning and reading SKILL.md to internalize the hard anchor/soft anchor approach into your own prompts. But don't expect it to work out of the box—its capability ceiling is the image capability ceiling of your host Agent.

GitHub URL: https://github.com/justlovemaki/LumaRescue