Turn the process into a world that can be followed along as one watches, and deliver a real image, not just a prompt. First make sure the story can be read, then let materials and lighting add appeal.
Applicable Boundaries
- Suitable when the user has already given, or can clearly agree on, a 3–5 step linear process. The number of steps is not fixed at five.
- When there are parallel, loop-back, or exception paths, do not quietly cut them; first explain which main line this image shows. When the user needs the complete logic, switch to a flowchart recommendation instead, and do not pass it off as a precise engineering diagram.
- Distinguish functional analogies from real mechanical processes. Software cores being packaged can be a metaphor; coffee cannot flow directly from whole beans into a beverage.
- This is an artistic image with process semantics, and it does not promise that one wordless image can uniquely reverse-engineer abstract professional terminology.
- Use only references authorized by the user. Do not treat one prototype's track, color scheme, materials, or fixed characters as a universal template.
Input and Capability Precheck
First use the minimum necessary judgment to decide whether to start drawing, so that the story is not quietly altered just to fit the frame:
- Only a subject, no process: you may propose a 3–5 step main line explicitly labeled as a demonstration, and explain the key assumptions. Do not present steps you added yourself as user-confirmed or industry-standard.
- More than five steps: only when the user allows summarization may adjacent sub-actions be merged; preserve the complete mapping from the original steps to the macro stages and the original order. If each step must remain independently visible, do not arbitrarily cut steps or force them into five stations; explain the limitation, propose splitting into multiple complete worlds or switching to a flowchart, and wait for the user to choose.
- When branches/loops must be fully expressed: explain that a single-path artistic image cannot faithfully carry all the logic, and first give a flowchart or a proposed single main line to be confirmed; without authorization, do not discard paths such as failure, refund, or retry.
- When the user only wants a plan or explicitly forbids image generation, deliver the plan and mark it as "not generated, visual not verified"; do not call an image generation tool, and do not create final.png.
- When image generation tools are missing: you may deliver an executable prompt to be run, but clearly state that this is a degraded delivery, not a finished product. When image-viewing capability is missing, do not claim visual acceptance passed; when an independent reviewer is missing, record self-review and independent review separately.
For an already clear simple process, proceed directly, and do not turn the precheck into a long form for the user to fill out. For capabilities and trial access, see references/runtime.md。
First Make a Short Contract
Record these before generation, without asking the user to fill out a long form:
- The known process, the macro number of steps, and the reading direction.
- Each step's input state → single main action → output state.
- The identifying features of the carrier throughout, and the form changes that can legitimately occur.
- How one continuous path passes through each step, and where the start and end points are.
- How adjacent actions are distinguished; which props are only background and should not be read as an extra step.
- Aspect ratio, text requirements, user constraints, and the generation limit for this round.
When no process is provided and the user only says "give it a try," choose a common-sense scenario close to the conversation, and make clear that this is a chosen demonstration main line, not the only industry standard.
Make the macro nodes clear, rather than dividing the image into boxes
- Keep only one primary action per node. Separate nodes with spatial whitespace, terrain transitions, and distinct silhouettes, not with frame lines or isolated islands.
- Sub-actions belonging to the same stage should be completed at the same workbench, building, or actor group, avoiding packaging, labeling, and outbound shipping being unintentionally expanded into three stations.
- Connecting lines can be roads, conveyor belts, narrow channels, desktop operation strips, terrain, or the motion traces of a carrier; there is no need to draw an orange S-shaped track every time.
- Paths should be traceable and should not form misleading forks, dead ends, or self-intersections. Only brief occlusion is allowed, and the entrance-exit relationship must be visible.
- The starting point shows the input, and the endpoint shows the use after completion. The endpoint can be a large main scene, but do not attach another independent ending that looks like a sixth station.
- A static image can repeat different time states of the same carrier, making it clear that this is a compression of continuous time, not multiple different items. Maintain identity with color, shape, or a single pattern; when the process changes, retain a reasonable association, and do not force coffee powder to still grow into coffee beans.
- Adjacent actions use different evidence: automated testing uses an unmanned probe, human review uses two people comparing; grinding uses beans entering and powder coming out, and brewing uses water passing through the coffee bed.
- Props and characters can naturally cross regions, preserving the design feel of one complete world.
Art direction serves the subject matter
First determine materials, lighting, color palette, scale, and visual rhythm, then form a short prompt. Do not equate "refined" with metal gears everywhere, glowing cores, and orange-cyan color schemes.
Optional directions, not mandatory templates:
- Precision industry: glass, anodized metal, ceramics, a small amount of work lighting.
- Paper and wood workshop: wood grain, paper fibers, pottery, fabric, and diffused daylight.
- Living neighborhood: stucco buildings, cardboard boxes, asphalt, plants, and natural human activity.
- Natural processing: soil, water, plants, and material changes genuinely visible in the process.
There is detail near the primary action, and the connections are quiet. Small characters do not mean unrecognizable; avoid depth of field blurring the entire latter half of the process. When references are needed, first observe their materials and spatial relationships, and do not copy specific compositions or proprietary logos.
Text is a controlled option
- When the user requests no text, generate a clean image and deliver a brief Chinese process description; if the professional semantics cannot be uniquely identified, say so directly.
- When the user has not requested no text and the subject matter is abstract, you may use one 2–4 character environmental sign per station, adding a sequence number and a short title when necessary. The Chinese must match the contract word for word, and the number of signs must not expand.
- Signs only confirm steps and must not be used to compensate for missing actions with large blocks of copy. Do not split the image into a grid to add text.
- If generated text is wrong, revise only the incorrect characters and recheck the entire image; when text is moved back outside the image for explanation, inform the user, and do not treat incorrect text as the final image.
Generation and revision
Use the image generation tools available in the current environment, generating the entire page at once. Set tool or model parameters according to actual available capabilities, and do not fake seed determinism, quality levels, or image file paths.
- Write the prompt in the order of "world and materials → path → sequential actions → carrier changes → text/constraints".
- The first round defaults to one main image; after self-review, at most two targeted revisions, unless the user provides another budget.
- Keep old drafts and clearly write the reasons for selection and elimination; when revising, lock the composition, path, characters, and stages that have no problems.
- Do not generate each station separately and then stitch them together. Native text or environment and the subject are generated as one integrated whole; if the user explicitly wants editable annotations, a guide file can be provided separately.
- If there are still important semantic errors after two consecutive revisions, report that the standard was not met; do not end by slapping on a "verified" label.
Finished-product QA
Use references/qa.md to check the final file; do not review only the prompt. At minimum, look at the original size and a proportional thumbnail.
Before delivery, run the zero-dependency file validator:python3 scripts/verify_delivery.py <交付目录> --ratio 3:4 --require-all-docs(replace the aspect ratio with the actual agreed one). If there is an independent review JSON, add --review <review.json> to verify whether its reviewed_image_sha256 points to the final image. This script only checks PNG integrity, aspect ratio, the four-piece documentation set, and hash binding; the output ok absolutely does not mean the image or process semantics passed.
Judge along two axes: "visual completeness" and "process fidelity." The following are release blockers:
- core steps missing, out of order, or violating the agreement;
- path broken or misleading branches produced;
- carrier changes without causation, key hand-and-object/water-and-container relationships not holding;
- garbled text, text mismatched with steps, main figures or objects accidentally cut off;
- what was actually generated is a grid collage rather than a complete scene.
Inability to blind-guess technical terms without text is not an automatic failure; if the image is left with only decoration and any process could be explained by changing the title, process fidelity is still unqualified. The thumbnail should retain the path, macro nodes, and main carriers; it is not required to make every tiny component legible.
Delivery
By default, show the user the final image and a one-sentence sequence description. Save:
final.png: the final selected complete finished product;prompt.md: the generation and revision prompts actually used;scene-contract.md: short contract;qa.md: QA bound to the final file hash, including failed and unverified items.
Thumbnails, alternate drafts, and execution records go in the evaluation workspace. Research-stage baselines and private reference ledgers are not packaged with the public Skill. Saving files does not mean authorization to release, upload, install, or push has been obtained.