[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6K7JPHPXHMTQCGA08V42KRQ":3,"skill-same:01M6K7JPHPXHMTQCGA08V42KRQ":81},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":40,"linkOnly":41,"markdownContent":42,"markdownBody":43,"frontMatter":44,"refName":46,"commitSha":47,"skillDirectory":48,"forksCount":14,"repositoryHtmlUrl":49,"repositoryHomepageUrl":6,"ownerHtmlUrl":50,"ownerBlogUrl":6,"ownerDescription":51,"updatedAt":52,"install":53,"resources":66},"01M6K7JPHPXHMTQCGA08V42KRQ","delegate-image",null,"Generate a visual — SVG via text models (Astra) for logos and diagrams, a raster panel for photos — judged by two independent models. Use for any image ask.","joewalnes\u002Fskills","joewalnes","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F48950?v=4","01M6K7J46C00DDV34GMGT2NK72","Joe Walnes","01M6K7J4ANFVAP1P9ZJ3K4F3P2",3,"MIT","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fdelegate-image\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000016","图像与多媒体","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39],"image","delegate","delegate image","generate","text","ask","use","diagrams","models","photos","what the","visual","SVG","via","panel","for","2026-11-10 10:12:30",false,"---\nname: delegate-image\ndescription: Generate a visual — SVG via text models (Astra) for logos and diagrams, a raster panel for photos — judged by two independent models. Use for any image ask.\nargument-hint: \u003Cwhat the visual should be>\n---\n\n# Delegate: visuals — pick the output type, then the panel\n\nClaude can't generate raster images, but it can write SVG, and so can any capable text model. So the first decision is not which model — it's **which kind of file the user actually wants.** Get that wrong and a photorealistic model produces a blurry logo, or a vector model is asked for a sunset.\n\n## 1. Choose the output type\n\n| The ask | Output | Why |\n|---|---|---|\n| Logo, mark, icon, badge, diagram, chart, schematic, pictogram — anything with flat shapes, text, or that must scale, be edited, or **animate** | **SVG** | Resolution-independent, editable, tiny, animatable with CSS or SMIL, and reviewable as source. The user will want to change a colour or a word later; with a PNG they can't. |\n| Photo, scene, person, product shot, texture, illustration with lighting and depth — anything photorealistic or painterly | **Raster** (PNG) | This is what image-generation models are for; no text model can do it. |\n| \"Animated logo\", \"loading spinner\", \"hero animation\" | **SVG with animation** | CSS `@keyframes` inside `\u003Cstyle>` (broadest support) or SMIL; the static first frame must stand on its own. |\n\nWhen in doubt: *if the user would ever want to edit it, it's SVG.* Say which type you chose and why in one line; if the ask genuinely straddles (a logo *and* a photo of it on a mug), do both.\n\n## 2a. SVG panel — text models\n\n```bash\npython3 \u003Cskill-dir>\u002Fscripts\u002Fsvg-panel.py \"\u003Cprompt>\" \u003Coutput-dir> [--models openai\u002Fgpt-6-astra,z-ai\u002Fglm-5.3]\npython3 \u003Cskill-dir>\u002Fscripts\u002Fsvg-panel.py --render-only \u003Coutput-dir>\u002Fclaude.svg\n```\n\n- **Models:** `openai\u002Fgpt-6-astra` (ZDR on OpenRouter, ~$0.10 per graphic at $10\u002F$50 per Mtok — the strongest text model available for this) and `z-ai\u002Fglm-5.3`. Astra is used **only here** — it has no image output, so it never appears in the raster panel.\n- **Claude's own entry.** The Agent running this also writes one itself as `\u003Coutput-dir>\u002Fclaude.svg`, then renders it with `--render-only`. Three entries, three lineages.\n- The script asks each model for *only* an SVG, extracts the first `\u003Csvg>…\u003C\u002Fsvg>`, rejects anything that doesn't parse as XML, saves `\u003Cmodel>.svg`, and renders `\u003Cmodel>.png` with `rsvg-convert` so the judges can *see* it. Animated SVGs render as their first frame.\n- **Write the prompt for a vector artist**: shapes and their relationship, palette as hex, \"no text\" unless wanted, target sizes (\"works at 32px\"), and for animation: what moves, how long, and that the first frame must be presentable.\n\n## 2b. Raster panel — image models\n\n```bash\npython3 \u003Cskill-dir>\u002Fscripts\u002Fimage-panel.py \"\u003Cdetailed prompt>\" \u003Coutput-dir>\n```\n\nRuns three Gemini image tiers in parallel (the only ZDR-clean image generators on OpenRouter — OpenAI's image models have no ZDR endpoint), saves `flash.png`, `nano-banana.png`, `pro.png`, and writes per-model cost. Benchmarked 2026-09-01 on 4 prompts, blind-judged:\n\n| Rank | Model | Notes |\n|---|---|---|\n| 1 | `google\u002Fgemini-3.1-flash-image` | Won 3 of 4; best all-rounder |\n| 2 | `google\u002Fgemini-2.5-flash-image` (\"nano banana\") | Won the photorealistic nature shot |\n| 3 | `google\u002Fgemini-3-pro-image` | Won the logo test; priciest (~$0.14\u002Fimg) — but logos are SVG now |\n\nThe user generates images rarely, so the 3-model panel (~$0.24) is the default; don't downgrade to one model to save money unless asked. To edit an existing raster, call `scripts\u002Fgenerate-image.py` directly with `-i input.png`.\n\n## 3. Two independent judges — same for both panels\n\nSequentially (concurrent calls can trip OpenRouter's credit hold), two vision models from lineages other than the generators':\n\n```bash\ncd \u003Coutput-dir> && pi -p --no-tools --no-session --provider openrouter --model moonshotai\u002Fkimi-k2.5 -- @a.png @b.png @c.png \"\u003Cjudge prompt>\"\ncd \u003Coutput-dir> && pi -p --no-tools --no-session --provider openrouter --model z-ai\u002Fglm-4.6v      -- @a.png @b.png @c.png \"\u003Cjudge prompt>\"\n```\n\nFor **raster**: \"These N images were generated from the prompt '\u003Cprompt>'. For each, a one-sentence critique; then your favourite with a one-sentence reason.\"\n\nFor **SVG**, attach the PNGs *and* paste the SVG sources, and ask additionally: does it read as the concept at 32px and at 256px; is the source clean (a `viewBox`, no embedded rasters, no scripts, sane path count); for animation, does the described motion serve the mark or decorate it. A judge that can't see images will still produce a plausible critique from filenames — confirm `pi --list-models \u003Cname>` shows `images: yes`, and that the critique cites visual specifics.\n\n## 4. Execution — one Agent call, one clean response\n\nThe user invokes this and expects **the files plus two quick critiques in a single response**, not a play-by-play. All of it — the type decision, the panel, Claude's own SVG, rendering, judging — happens inside **one Agent tool call** (general-purpose). It runs asynchronously; after dispatching, say nothing and wait for the completion notification. Brief the agent with the ask, the type you chose, both command lines, the judge prompts, and: send every output with SendUserFile (`display: render`; for SVG send both the `.svg` and its `.png`), then return **only** each judge's pick and one-sentence reason, your own one-sentence recommendation, and cost.\n\nIf any raster is a photorealistic depiction of a real, identifiable person, say so in one line and keep it out of anything published.\n\n## 5. After the Agent returns\n\nRelay in one short message: which entry each judge picked and why, your own take, the cost. Nothing about mechanics.\n","# Delegate: visuals — pick the output type, then the panel\n\nClaude can't generate raster images, but it can write SVG, and so can any capable text model. So the first decision is not which model — it's **which kind of file the user actually wants.** Get that wrong and a photorealistic model produces a blurry logo, or a vector model is asked for a sunset.\n\n## 1. Choose the output type\n\n| The ask | Output | Why |\n|---|---|---|\n| Logo, mark, icon, badge, diagram, chart, schematic, pictogram — anything with flat shapes, text, or that must scale, be edited, or **animate** | **SVG** | Resolution-independent, editable, tiny, animatable with CSS or SMIL, and reviewable as source. The user will want to change a colour or a word later; with a PNG they can't. |\n| Photo, scene, person, product shot, texture, illustration with lighting and depth — anything photorealistic or painterly | **Raster** (PNG) | This is what image-generation models are for; no text model can do it. |\n| \"Animated logo\", \"loading spinner\", \"hero animation\" | **SVG with animation** | CSS `@keyframes` inside `\u003Cstyle>` (broadest support) or SMIL; the static first frame must stand on its own. |\n\nWhen in doubt: *if the user would ever want to edit it, it's SVG.* Say which type you chose and why in one line; if the ask genuinely straddles (a logo *and* a photo of it on a mug), do both.\n\n## 2a. SVG panel — text models\n\n```bash\npython3 \u003Cskill-dir>\u002Fscripts\u002Fsvg-panel.py \"\u003Cprompt>\" \u003Coutput-dir> [--models openai\u002Fgpt-6-astra,z-ai\u002Fglm-5.3]\npython3 \u003Cskill-dir>\u002Fscripts\u002Fsvg-panel.py --render-only \u003Coutput-dir>\u002Fclaude.svg\n```\n\n- **Models:** `openai\u002Fgpt-6-astra` (ZDR on OpenRouter, ~$0.10 per graphic at $10\u002F$50 per Mtok — the strongest text model available for this) and `z-ai\u002Fglm-5.3`. Astra is used **only here** — it has no image output, so it never appears in the raster panel.\n- **Claude's own entry.** The Agent running this also writes one itself as `\u003Coutput-dir>\u002Fclaude.svg`, then renders it with `--render-only`. Three entries, three lineages.\n- The script asks each model for *only* an SVG, extracts the first `\u003Csvg>…\u003C\u002Fsvg>`, rejects anything that doesn't parse as XML, saves `\u003Cmodel>.svg`, and renders `\u003Cmodel>.png` with `rsvg-convert` so the judges can *see* it. Animated SVGs render as their first frame.\n- **Write the prompt for a vector artist**: shapes and their relationship, palette as hex, \"no text\" unless wanted, target sizes (\"works at 32px\"), and for animation: what moves, how long, and that the first frame must be presentable.\n\n## 2b. Raster panel — image models\n\n```bash\npython3 \u003Cskill-dir>\u002Fscripts\u002Fimage-panel.py \"\u003Cdetailed prompt>\" \u003Coutput-dir>\n```\n\nRuns three Gemini image tiers in parallel (the only ZDR-clean image generators on OpenRouter — OpenAI's image models have no ZDR endpoint), saves `flash.png`, `nano-banana.png`, `pro.png`, and writes per-model cost. Benchmarked 2026-09-01 on 4 prompts, blind-judged:\n\n| Rank | Model | Notes |\n|---|---|---|\n| 1 | `google\u002Fgemini-3.1-flash-image` | Won 3 of 4; best all-rounder |\n| 2 | `google\u002Fgemini-2.5-flash-image` (\"nano banana\") | Won the photorealistic nature shot |\n| 3 | `google\u002Fgemini-3-pro-image` | Won the logo test; priciest (~$0.14\u002Fimg) — but logos are SVG now |\n\nThe user generates images rarely, so the 3-model panel (~$0.24) is the default; don't downgrade to one model to save money unless asked. To edit an existing raster, call `scripts\u002Fgenerate-image.py` directly with `-i input.png`.\n\n## 3. Two independent judges — same for both panels\n\nSequentially (concurrent calls can trip OpenRouter's credit hold), two vision models from lineages other than the generators':\n\n```bash\ncd \u003Coutput-dir> && pi -p --no-tools --no-session --provider openrouter --model moonshotai\u002Fkimi-k2.5 -- @a.png @b.png @c.png \"\u003Cjudge prompt>\"\ncd \u003Coutput-dir> && pi -p --no-tools --no-session --provider openrouter --model z-ai\u002Fglm-4.6v      -- @a.png @b.png @c.png \"\u003Cjudge prompt>\"\n```\n\nFor **raster**: \"These N images were generated from the prompt '\u003Cprompt>'. For each, a one-sentence critique; then your favourite with a one-sentence reason.\"\n\nFor **SVG**, attach the PNGs *and* paste the SVG sources, and ask additionally: does it read as the concept at 32px and at 256px; is the source clean (a `viewBox`, no embedded rasters, no scripts, sane path count); for animation, does the described motion serve the mark or decorate it. A judge that can't see images will still produce a plausible critique from filenames — confirm `pi --list-models \u003Cname>` shows `images: yes`, and that the critique cites visual specifics.\n\n## 4. Execution — one Agent call, one clean response\n\nThe user invokes this and expects **the files plus two quick critiques in a single response**, not a play-by-play. All of it — the type decision, the panel, Claude's own SVG, rendering, judging — happens inside **one Agent tool call** (general-purpose). It runs asynchronously; after dispatching, say nothing and wait for the completion notification. Brief the agent with the ask, the type you chose, both command lines, the judge prompts, and: send every output with SendUserFile (`display: render`; for SVG send both the `.svg` and its `.png`), then return **only** each judge's pick and one-sentence reason, your own one-sentence recommendation, and cost.\n\nIf any raster is a photorealistic depiction of a real, identifiable person, say so in one line and keep it out of anything published.\n\n## 5. After the Agent returns\n\nRelay in one short message: which entry each judge picked and why, your own take, the cost. Nothing about mechanics.\n",{"name":5,"description":7,"argument-hint":45},"\u003Cwhat the visual should be>","main","0d32bda2cb8e014372c2814f5aa47f48eee0f411","skills\u002Fdelegate-image","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills","https:\u002F\u002Fgithub.com\u002Fjoewalnes","Experimental Projects @ Stripe.\r\n\r\nPreviously: Apple, Google, ThoughtWorks, DRW, Startups.","2026-09-09 00:56:55",{"source":8,"skillName":5,"commands":54,"prompt":64,"zipAvailable":65},[55,58,61],{"tool":56,"command":57},"npx","npx skills add joewalnes\u002Fskills --skill delegate-image",{"tool":59,"command":60},"bunx","bunx skills add joewalnes\u002Fskills --skill delegate-image",{"tool":62,"command":63},"pnpm","pnpm dlx skills add joewalnes\u002Fskills --skill delegate-image","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fdelegate-image\u002FSKILL.md",true,[67,73,77],{"path":68,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":71,"sourceHtmlUrl":72},"scripts\u002Fgenerate-image.py","SCRIPT","脚本",3116,"https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fdelegate-image\u002Fscripts\u002Fgenerate-image.py",{"path":74,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":75,"sourceHtmlUrl":76},"scripts\u002Fimage-panel.py",1543,"https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fdelegate-image\u002Fscripts\u002Fimage-panel.py",{"path":78,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":79,"sourceHtmlUrl":80},"scripts\u002Fsvg-panel.py",4358,"https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fdelegate-image\u002Fscripts\u002Fsvg-panel.py",[82,98,110,128,144,150],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":40},"01M6K7MSG99DXYKTVVFGZWND1S","tool-web","Build a lightweight single-file web application with no external dependencies","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Ftool-web\u002FSKILL.md",[],[89,90,91,92,93,94,95,96,97],"tool web","tool","web","file","single-file","build","dependencies","application","with",{"skillNo":99,"name":100,"nameZh":6,"description":101,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":102,"categories":103,"keywords":104,"publishedAt":40},"01M6K7MPN2FG6M6JQFWFR18CJ0","todo","Add a new bug or todo to the project's tracker","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Ftodo\u002FSKILL.md",[],[100,105,106,107,108,109],"bug","tracker","task","project","add",{"skillNo":111,"name":112,"nameZh":6,"description":113,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":114,"categories":115,"keywords":116,"publishedAt":40},"01M6K7MGQFFYCD8HFDP1B95KDN","slop","Find unchosen code — surface area growing without leverage. Use pre-merge on a branch, or to check if a repo is accreting.","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fslop\u002FSKILL.md",[],[112,117,118,119,120,121,122,123,124,125,30,126,127],"pre-merge","rev","explain","merge","code","find","branch","repo","check","leverage","quick",{"skillNo":129,"name":130,"nameZh":6,"description":131,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":132,"categories":133,"keywords":134,"publishedAt":40},"01M6K7MCVREA4NZ46Q9JWZJ7TG","sitrep","Where were we? Branch, in-progress work, blocked agents, inbox, next steps. Use when resuming a session.","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fsitrep\u002FSKILL.md",[],[130,135,136,137,138,139,30,140,141,142,123,143],"work","grep","Bash","inbox","session","next","progress","next steps","read",{"skillNo":145,"name":146,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":147,"categories":148,"keywords":149,"publishedAt":40},"01M6K7M99FYMH9SFSEKB718J2B","scorecard","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fscorecard\u002FSKILL.md",[],[146],{"skillNo":151,"name":152,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":153,"categories":154,"keywords":155,"publishedAt":40},"01M6K7M4HNZR34GSB3V14KVCWA","retro","https:\u002F\u002Fgithub.com\u002Fjoewalnes\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fretro\u002FSKILL.md",[],[152]]