CommunityImagegithub.com

liusucheng1999-lab/imagegen-poster-workflow

Design, generate, redesign, or locally edit posters, brand share graphics, covers, and promotional images using built-in ImageGen. Ideal for ChatGPT Work or Codex tasks that require full visual output, brand reference and asset protection, step-by-step review, and iterative revisions.

What is imagegen-poster-workflow?

This skill provides a structured, professional workflow for creating and iterating on promotional posters and visual assets using the built-in ImageGen tool (available in ChatGPT Work, Codex, and similar AI agent platforms). It is designed for complex design tasks where maintaining brand identity, protecting original assets (like QR codes and logos), and achieving pixel-perfect final outputs are critical. The workflow defines three core modes: **Generate** (for new designs from scratch), **Redesign** (for complete visual overhauls of existing posters while locking brand assets), and **Edit** (for targeted local modifications to a specific version). It enforces a clear separation between confirmed requirements and design assumptions, mandates the tracing of all assets (reference images, source materials, and edit baselines), and provides a rigorous acceptance checklist covering both content/functionality and overall design quality. The skill also includes detailed instructions for handling multilingual text, functional QR codes, and brand assets, as well as a systematic process for handling negative feedback by verifying assumptions before proceeding. Under the hood, the skill guides the agent to compose detailed image generation prompts that explicitly describe the communication task, visual hierarchy, asset roles, and protection requirements. It prioritizes using the built-in ImageGen tool directly for a seamless, all-in-one output, avoiding unnecessary external rendering or HTML/CSS approaches. It also mandates keeping a complete version history with prompts, baselines, and acceptance status for traceability and iterative refinement.

Works with~Claude Code✓Codex CLI~Cursor
npx skills add https://github.com/liusucheng1999-lab/imagegen-poster-workflow/tree/HEAD

Installed? Explore more Image skills: steipete/songsee, affaan-m/frontend-design, mattpocock/improve-codebase-architecture · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Poster Generation Workflow

Requirements and Modes

Extract from the current request, confirmed context, and related project materials: the purpose, placement, the viewer's next action, canvas dimensions, exact copy, necessary functionality, locked assets, and prohibited items. For continuation work, inherit still-applicable confirmed requirements. First, consult existing materials and assets; then, ask for truly missing key items. Decide on conventional design details autonomously.

Use a brief checklist to record delivery essentials and their sources, separating confirmed requirements from current design assumptions. If the purpose includes scanning to enter a product, locate the actual functional code before the first output and incorporate the scan entry point into the layout. Whether a code is needed depends on the purpose and context. When essential materials are missing, state the specific missing item, retain delivery requirements; only deliver a preview with clear status when the user wants to see a direction or accept a preview.

ModeWhen to UseScope of Change
GenerateDesigning from scratch, or providing only style, content, etc., as referenceFreely conceive the entire poster within requirements and brand constraints
RedesignUser wants a visual overhaul based on an existing posterLock brand identity, core information, and specified assets; can redo composition, hierarchy, and art style
EditUser requests specific changes to a particular versionStart from the user-selected or most recently approved version; modify only the target and necessary adjacent areas, protecting other parts

These three modes describe the design intent. Select the appropriate generation, reference, or editing approach supported by the current built-in ImageGen for the task.

Reference Images and Source Files

For each image, clarify its role, the content to borrow or lock, the variable range, and details to be discarded. One image can serve multiple roles; describe by area.

  • Style references capture lighting, color, medium, or texture. Composition references capture visual focal point, eye path, or layout relationships. Content references capture factual scenes or objects. Exclude irrelevant text, codes, and patterns from reference images.
  • Brand or character references lock features that identify identity, such as silhouette, proportions, color scheme, logo, and product appearance. Edit baselines are the benchmarks for targeted modifications and acceptance comparison.
  • Assets that must be used as-is retain independent source files, including functional codes, official logos, finalized text graphics, or specified product assets. Preserve unmodified originals and baselines; output each generation or edit to a new version.
  • Aesthetic references prioritize works previously approved by the user or explicitly provided design references; record the specific characteristics to be achieved this time. Character images constrain only character identity, and alone cannot justify the poster's overall style choice. Historical images within the project must have their usage and approval status verified; rejected drafts are used to record unsuccessful directions.
  • In case of conflict, prioritize satisfying the user's current request and explicit locking items. Second, adhere to confirmed brand constraints. Finally, borrow from style or composition references. Actually pass relevant reference images and edit baselines to ImageGen, and specify each image's role in the prompt. If key baselines or images cannot be passed, ask the user to reattach them.

Creative Process and Image Generation

  1. First, determine the communication task this design aims to solve and the rationale for the chosen direction. Describe in one or two sentences how the visual main line, information priority, and functional entry points collectively serve the task. The style choice should relate to the purpose or aesthetic references. If the aesthetic basis is insufficient, mark the direction as a hypothesis to be verified and proceed with comparable trial drafts.
  2. When the direction is clear, generate directly. When there are significant directional differences for Generate or Redesign, first propose a small number of truly distinct candidates in concept or visual organization, filter using purpose and aesthetic references, and briefly note the reasons. If the textual concept is still insufficient for judgment, prioritize generating two comparable thumbnails, proceeding with the user's specified quantity or scope. Retain necessary copy, functional areas, and brand constraints for each, ensuring the comparison is on overall design. For Edit, continue with the approved direction and clarify the local objective.
  3. In the prompt, first write the task and exact copy, then the reference roles, protection items, and a brief creative main line and information hierarchy. Let ImageGen coordinate the specific composition, relationship between text and graphics, lighting, and materials. Only specify precise positions, ratios, and coordinates when needed for user specifications, reference alignment, local modifications, or embedding real assets. Avoid locking down all unverified assumptions about fonts, colors, poses, textures, and layouts. Include only text confirmed by the user or task-required visible text; retain all mandatory information.
  4. Before calling the tool, systematically compare the delivery checklist against the actual prompt sent. Check for omissions, unauthorized downgrading to trial drafts, writing essential items into prohibited items, and whether reference images have been actually passed. When embedding real assets in a subsequent step, still incorporate their space and hierarchy into the prompt and record the subsequent embedding steps. If conflicts are found, correct the prompt first.
  5. Call the built-in ImageGen within the current session, allowing titles, illustrations, layout, and screen elements to be generated together, resulting in a complete, independently valid poster. When real assets need to be embedded as-is, arrange corresponding areas within the composition; the rest of the layout and visuals are still completed by ImageGen. Avoid defaulting to generating a text-free base image or rendering the entire poster with HTML/CSS, SVG, or Python. If the built-in tool is unavailable, state this fact; only proceed with a path requiring an API key or additional charges if explicitly authorized by the user.

Text and Functional Code Protection

  • Text: Adhere strictly to the user's provided original text. Preserve every character for simplified/traditional Chinese, numbers, dates, punctuation, English case, and proper nouns. Generate regular copy with ImageGen and verify. If pixel-perfect preservation of existing text assets is needed, use the original asset directly. If key copy remains incorrect after targeted edits and delivery requires word-for-word accuracy, use the confirmed original text for precise typesetting in the corresponding area.
  • QR codes, Mini Program codes, and official logos: Use the original files provided or confirmed by the user. For official drafts, functional codes must always originate from real assets. When permitted by user requirements and current tool rules, embed them only in the corresponding area using deterministic tools, preserving proportion and clarity. Retain the complete pattern, quiet zone, and contrast of the functional code; avoid covering, deforming, cropping, or applying filters. After embedding, re-inspect the overall composition, hierarchy, and relationship of the scan area within the final artwork. After output, verify identification using the corresponding entry point at the final publish size; when the target is known, also verify correctness. Verify Mini Program codes with WeChat. Mark verification status separately for items dependent on actual business results (e.g., attribution).
  • If original functional codes are missing, ask the user to provide them. For preview drafts, functional areas can be left blank; avoid drawing codes that appear usable but are not. When actual scanning verification is not possible, mark the delivery note with "Pending Live Scan" and retain it as a task for release acceptance. If the user requires pure ImageGen and prohibits compositing, adhere to the restriction and explain the items pending verification (e.g., precise text, pixel-level protection, or live scanning).

Acceptance and Targeted Modifications

  • Content and Functional Acceptance: Check against the delivery checklist: copy, dimensions, brand identity, original assets, functional entry points, and prohibited items. Inspect at original size for cropping, typography, contrast, and artifacts. Missing items, typos, changes to locked assets, or failed live scans must be corrected. List incomplete check items with their status.
  • Overall Design Acceptance: View the full image and its actual placement thumbnail alongside aesthetic references or candidates. Judge whether the visual main line is established, the information hierarchy is clear, the text, graphics, and functional areas form a cohesive whole, and the style completeness matches the references. Briefly note the specific visual rationale for selecting or requesting a rework. When there is only one draft and aesthetic references are missing, keep the aesthetic judgment tentative; produce a truly different candidate if necessary. Correct text, pixel consistency, and "visually inspected" cannot substitute for overall design judgment.
  • Record Conclusions Separately: Content/functional, overall design, and user approval should each be recorded independently. Only submit as a final candidate when the first two have both met the current requirements. Tool verification results only support their corresponding items. If the user has not yet approved, retain a "Pending Feedback" status; after negative feedback, revoke that version's design approval conclusion.
  • Trace Back Negative Feedback First: When receiving feedback like "doesn't look good" or "wrong direction," first trace back the understanding of the purpose, aesthetic references, usage scope of role assets, and basis for selecting candidates before determining new hypotheses for the next round. If there is an already approved direction and a clear local issue, proceed with Edit. If the overall direction is not established, proceed with Redesign, changing the concept or visual organization, avoiding just adding layout instructions or swapping decorations. When supplementary aesthetic basis is needed, reuse confirmed references or use clearly different trial drafts for the user to compare.
  • Targeted Modifications and Stopping: Start local modifications from the version selected by the user or closest to the requirements. Focus each round on fixing one issue or a set of root-related issues. After completion, check the target and unaffected areas against the baseline, then redo the two acceptance checks. If the same hypothesis shows no improvement after retrying, go back upstream to change direction or report the specific blocking point. Respect the user's scope limitations; avoid continuous re-rolling without basis.

Prompts, Versioning, and Delivery

  • Preserve the user's original requirements, delivery essentials and sources, and the complete prompt sent in each round. Additionally, record the mode, baseline version, reference images and original assets actually passed, direction rationale, modification points, and separately recorded acceptance status. When negative feedback occurs, note the rejected hypothesis and the change to be verified in the next round. Prompts are live records; post-hoc summaries should be noted separately.
  • Use the same project name with incremental versions (e.g., poster-v01.png, poster-v02.png), recording the corresponding prompt and parent version for each version. Preserve original assets, the user-selected baseline, and the final approved draft. When continuing edits, start from a clearly identified version. If original assets have been embedded, differentiate between the ImageGen raw output and the delivery-ready final version, and record the embedded items.
  • Save the final deliverable to the location specified by the user or as per project conventions. When a file needs to be saved, store the prompt and version records alongside the corresponding artwork, or use the project's existing recording method. For in-session previews only, delivery images and actual prompts can remain within the conversation. Default to delivering the finalized version that has passed acceptance, and explicitly list items still requiring user verification.

Related Skills