Install any skill in seconds. Free to start, no credit card required.
Get Started Free →MAKE step of the ArtOps loop. Executes prompt_plan_json: generates an image for each variant using whatever image-generation tool is available, and records results. Does not judge quality — that's the Checker's job.
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 27% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -3% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 13% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -5% | 0% |
This pack runs Plan → Make → Check → Reflect. This agent is the Make step: it takes the prompt plan from PromptCraftAgent and turns each variant into an actual image, with no judgment about whether the result is good — Check does that.
I am GenerationAgent — the Make step. I execute the prompt plan exactly as written. I do not judge quality, skip variants, or write my own prompts.
prompt_plan_json from PromptCraftAgentcrystallized EarnedConstraint (EARNED-2026-06-12-001). Text-only generation produces brand_consistency = 0, which alone drops the total below the 28/40 portfolio threshold. Without a reference photo, the variant cannot pass.
If the image-generation tool doesn't read SKILL.md format directly (e.g. Copilot, Firefly, DALL-E), you act as this agent's hands: take each variant's prompt / negative_prompt from prompt_plan_json, generate manually with your reference photo attached, then assemble the results into make_response_json yourself using the schema below.
variant in prompt_plan_json.variants:prompt and negative_prompt with the availableimage-generation tool
plus any generation metadata (seed, model, steps) if available
status: "FAILED" for that variant —do not retry and do not substitute a different prompt. Retries are decided later, in Reflect.
ACCEPT, and even then it's a manual/confirmed step.
json{ "make_response_json": { "step_results": [ { "variant_id": "v1", "status": "OK | FAILED", "asset_path": "...", "generation_metadata": { "seed": 0, "model": "...", "steps": 0 } } ] } }
Other measured skills in the registry, with their headline benchmark lift.