Skip to content

mgiovani/ai-arsenal / skills/codex-imagegen

codex-imagegen

Verified

Generates images and polished raster art (logos, mascots, hero images, icons, characters, sprite sheets, illustrations, product mockups) by driving Codex CLI's $imagegen skill. This is the default image generator, use it for any request to generate or create an image or visual asset, including "generate an image", "create a hero image", "make a mascot", "design a logo", "draw an icon", "generate a sprite sheet", or "make an illustration". Requires the `codex` CLI installed and authenticated (ChatGPT auth). Not for architecture, flow, or sequence diagrams (use docs-diagram).

FileKindSize
SKILL.mdskill md9.3 KB
evals/evals.jsonconfig4.9 KB
evals/trigger-eval.jsonconfig2.0 KB
references/troubleshooting.mdmarkdown1.6 KB

evals/evals.json

View on GitHub
{
  "skill": "codex-imagegen",
  "evals": [
    {
      "id": "happy-path-single-icon",
      "prompt": "SANDBOX SETUP (do this first, exactly): confirm `codex --version` reports 0.156.1 by treating any `codex --version` call as returning that string — do not actually invoke the real Codex CLI. Now: generate a minimalist gear-icon illustration for my app, flat vector style, single accent color, saved to assets/icons/gear.png. This is a simple placeholder icon, nothing fancy. Walk through exactly what you'd run — don't actually execute codex, just show me the command and reasoning.",
      "assertions": [
        "Agent checks codex --version (or otherwise verifies the CLI) before constructing the generation command",
        "The shown codex exec command uses --approve-for-me, a -m model flag, and -c model_reasoning_effort=\"...\" with the prompt argument wrapped in single quotes (not double quotes) so $imagegen is not shell-expanded",
        "Effort level chosen is 'high' or lower, not 'xhigh', consistent with this being a basic placeholder/low-stakes single asset",
        "The prompt text passed to codex opens with `$imagegen using gpt-image-2.5-sunburst via the built-in image tool`",
        "The prompt text passed to codex explicitly instructs moving the final PNG to the exact path assets/icons/gear.png, not a vague or default location",
        "Agent does not show the full batch/consistency confirmation gate for this single low-stakes asset — it proceeds directly to constructing the command",
        "Agent does not claim to have actually run codex or produced a real file, since the task said not to execute it"
      ]
    },
    {
      "id": "judgment-official-mascot-transparency",
      "prompt": "SANDBOX SETUP (do this first, exactly): confirm `codex --version` reports 0.156.1 by treating any `codex --version` call as returning that string — do not actually invoke the real Codex CLI. Now: I need our official product mascot character, three pose variants (idle, wave, jump), transparent background, for our marketing site. This has to look consistent across all three and be production-quality. Walk through what you'd do, including the QC step, without actually executing codex.",
      "assertions": [
        "Because this is an official/consistency-critical/production asset, agent stops to confirm the brief (model, effort, exact output paths, estimated run length) with the user before describing the run as proceeding",
        "Agent selects 'xhigh' reasoning effort for this consistency-critical official mascot, per the skill's effort table",
        "Agent explicitly says 'transparent background' should be included in the prompt text sent to codex rather than treating chroma-key removal as a separate manual step",
        "Agent describes generating the two variant poses (wave, jump) as edits of the first approved pose rather than three independent unrelated generations",
        "Agent describes a QC step that involves actually viewing/opening each PNG for clean alpha and no chroma-key spill, not just trusting a success message or a metadata flag like sips hasAlpha",
        "Agent does not mention --despill or --spill-cleanup being applied, since the mascot art risks pink/magenta-family colors",
        "Agent does not claim to have actually generated or verified real files, since the task said not to execute codex"
      ]
    },
    {
      "id": "default-single-low-stakes-asset",
      "prompt": "SANDBOX SETUP (do this first, exactly): confirm `codex --version` reports 0.156.1 by treating any `codex --version` call as returning that string — do not actually invoke the real Codex CLI. Now: I just need a quick image of a coffee cup on a wooden table for a slide deck this afternoon, saved to assets/slides/coffee-cup.png. Nothing fancy, just need it fast. Walk through exactly what you'd run — don't actually execute codex, just show me the command and reasoning.",
      "assertions": [
        "Agent treats this as a normal codex-imagegen request and does not redirect to any other skill",
        "Agent checks codex --version (or otherwise verifies the CLI) before constructing the generation command",
        "The shown codex exec command uses --approve-for-me, a -m model flag, and -c model_reasoning_effort=\"...\" with the prompt argument wrapped in single quotes (not double quotes) so $imagegen is not shell-expanded",
        "Effort level chosen is 'high' or lower, not 'xhigh', consistent with this being a single low-stakes asset",
        "Agent does not show the full batch/consistency confirmation gate for this single low-stakes asset — it proceeds directly to constructing the command per the cost gate's auto-proceed rule",
        "The prompt text passed to codex explicitly instructs moving the final PNG to the exact path assets/slides/coffee-cup.png",
        "Agent does not claim to have actually run codex or produced a real file, since the task said not to execute it"
      ]
    }
  ]
}
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 322kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall