redesign-concepts
Skill1 starsChanged 21 days agoMITView on GitHub
$redesign-concepts: built UI screenshots → image_gen → 10 redesigns → owner picks → build plan. Brief only → design-prototypes.
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 16
Use a fresh scratch dir inside the workspace that git ignores, for example .tmp/redesign-<n>/.
≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Redesign concepts
Critique an existing build from real screenshots and get generated redesigns of it. The images are design targets, not implementation: do not edit the product, commit images, or deploy. Default to 10 concepts; honor a different count. Not for exploring directions from a brief before anything is built.
Script paths are from the repository root; in a personal install, use the scripts/ folder in this skill's directory. Either way, run them from the workspace root, because every path they write is relative to it.
0. Preflight and cost
- Needs an image generation tool exposed in this session (built-in
image_gen) that accepts reference images, and a way to capture screenshots. Missing either: say so and stop; never draw concepts in code or call a paid API. - Each image costs about 60-70k tokens on the user's subscription, so 10 images is roughly 700k. Say so before starting the batch.
- Use a fresh scratch dir inside the workspace that git ignores, for example
.tmp/redesign-<n>/; the sandbox reads and writes only inside the workspace. - Load the project's own UI or art-direction skill or reference first, if it has one, and carry its constraints into the identity line.
1. Capture
Screenshot the current build at the size its users see (desktop 1440x900, or a phone viewport for a mobile-first product) with the project's verification setup and browser rules (AGENTS.md). Cover each key screen and state, especially the busiest: first paint, an open menu or drawer, a filtered or search state, an admin view, a detail page. Reuse captures only if fresh. Keep originals where the project or user keeps screenshots; copy the chosen references into <scratch>/refs/.
2. Critique, then concepts
Look at every capture and list the concrete problems this product actually has, for example repetition, too many controls, jargon, weak hierarchy, things pushed below the fold. Don't assume clutter: a sparse screen can hide what matters, and an expert tool may need more on screen. From that list, write the shared design direction and who the screens are for. Write the concepts: one screen, the problem it answers and one idea each, varied across composition, metaphor, hierarchy and interaction (recolors don't count). Mix user-facing and admin screens if both exist. Mark concepts that need data or backend the product lacks.
3. Prompts
Write <scratch>/concepts.json (format in the header of scripts/make-prompts.mjs), then run node .agents/skills/redesign-concepts/scripts/make-prompts.mjs <scratch>. It writes prompts/NN-name.txt, prompts/NN-name.refs (current screen first) and index.md. Every product-specific field is required and has no default: product, audience, identity (palette roles, type, card style from the screenshots), direction (goals all concepts share), orientation (landscape 1536x1024 or portrait 1024x1536), optional negatives, and per concept name, screen, problem, idea, refs, optional needs. Each prompt asks for a better version of the same screen with that problem, idea, direction, identity and the real data. Built in are only universal negatives: a near-copy of the screenshot, a different visual identity, device frames, placeholder or garbled text, more than one screen. Example: a light-themed admin app for non-technical staff used the direction "fewer controls visible at once, plain everyday words instead of jargon, one obvious primary action per area, clear grouping, progressive disclosure" and the negatives "dark mode", "a marketing hero", "emoji"; those fit that product only, so write your own.
4. Generate
For each prompt without a valid out/NN-name.png (a PNG of 100 KB or more), call the image tool with that prompt and its refs as input images, and save to the exact path. Skipping valid outputs lets a rerun resume. scripts/run-batch.sh <scratch> 3 drives separate codex exec processes three at a time instead; use it only when nested Codex runs are permitted here. A generated screen is hundreds of KB or more; a few-KB file was drawn some other way and does not count.
5. Review and present
Look at every image and note real flaws (invented labels, internal names in lists, garbled text, a copy of the screenshot). Regenerate only clear failures and say what changed. Show all images together with stable numbers, a one-line summary each, which need new data, and a recommended set with a concrete reason tied to the step 2 problems.
6. Fold picks into the build
On the owner's picks, write <scratch>/synthesis.md (copy it to the project's plans if the work outlives the session) resolving conflicts between concepts: one vocabulary across screens and the old terms it replaces, which existing features each idea replaces or hides, new data each needs and its source, and build order. Turn it into plan tasks or quality-loop checks. For a wow-loop or similar contract, add a versioned amendment with named checks and name each old check it retires because its feature is gone; never relax a check to make the new design pass.
In this repository85 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- addskillSkill
- adopt-repoSkill
- advocateSkill
- apply-firmwareSkill
- arenaSkill
- astra-reviewSkill
- babysit-ciSkill
- brainstormingSkill
- broSkill
- cache-loopSkill
- cavemanSkill
- claude-reviewSkill
- codex-reviewSkill
- dareSkill
- deep-planSkill
- design-prototypesSkill
- enhance-promptSkill
- external-reviewSkill
- fable-modeSkill
- forge-repo-ui-skillSkill
- handoff-auditSkill
- impartial-reviewSkill
- init-projectSkill
- labSkill
- long-horizonSkill
- optimize-contextSkill
- opus-fullreviewSkill
- perf-loopSkill
- recallSkill
- refineSkill
- serversSkill
- session-hubSkill
- showpieceSkill
- sync-starterSkill
- whySkill
- wow-loopSkill
- wrapupSkill
- writing-plansSkill
- writingSkill
- addskillSkill
- adopt-repoSkill
- advocateSkill
- apply-firmwareSkill
- arenaSkill
- astra-fullreviewSkill
- astra-reviewSkill
- babysit-ciSkill
- brainstormingSkill
- broSkill
- cache-loopSkill
- cavemanSkill
- claude-reviewSkill
- codex-fullreviewSkill
- codex-image-genSkill
- codex-reviewSkill
- coordinateSkill
- dareSkill
- deep-planSkill
- The first 60 shown.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

