--- name: hyper-imagegen description: "[Hyper] Use this skill when the user asks to create or edit one raster image, or to receive one ready-to-use image prompt. It compiles one brief, invokes the runtime's exposed image tool (in ChatGPT, gpt-image-2.6-sunburst when a model selector is exposed; otherwise the runtime default) rather than stopping at prompt delivery, and reports a verified terminal. Not for image analysis, reusable prompt packs, SVG/vector or code-native graphics, generic design documentation, or editable UI/code." compatibility: Requires an observed image-generation or image-editing capability. A runtime's exposed image tool (ChatGPT's native tool, or an agent or MCP image tool) may deliver its returned artifact directly; filesystem persistence additionally requires observed secure writing. Runs standalone in fileless runtimes; the bundled validator needs bash 3.2+ and jq 1.6+. Includes are package-relative. --- Open the rules as directed below. Resolve bundled file paths from the directory containing this document and project paths from the target project. - [rules/capability-and-output.md](rules/capability-and-output.md) - [rules/prompt-compilation.md](rules/prompt-compilation.md) - [references/runtime-capability-and-drift.md](references/runtime-capability-and-drift.md) - [references/prompt-crafting.md](references/prompt-crafting.md) - [references/visual-direction.md](references/visual-direction.md) # Hyper Imagegen > One request, one truthful artifact. Report in Korean by default, including the single question a route may ask and each terminal's evidence block. Preserve the user's exact rendered text, the compiled prompt, file paths, digests, and model names verbatim. Use another language only when the user asks for it. Turn a one-off create or edit request into a truthful delivered artifact, and handle a one-off prompt-only request the same way. Use this skill when the requested output is exactly one of: one new image, one edit of a supplied image, or one ready-to-use image prompt. Do not use it for image analysis or critique, reusable prompt packs and templates, generic design documentation, or editable UI and code. Do not use it when the asset is better made code-native: extending an existing SVG or vector icon set, logo system, or illustration library, or drawing simple shapes, diagrams, or wireframes that belong in SVG, HTML/CSS, or canvas. A per-request refusal of prompt saving is not a routing question; it is handled by the fallback policy in `rules/capability-and-output.md`. | Field | Contract | |---|---| | Intent | Deliver each requested image, edit, or one-off ready-to-use prompt faithfully. | | Trigger | A request for one image, one edit of a supplied image, or one image prompt, with the material needed to compile it or a route that may ask for it. | | Scope | Read supplied references as evidence; deliver runtime-returned artifacts directly, write files only under `.hyper/imagegen//` or, when the user names one, a new file at the named project destination; never modify an edit source or overwrite an existing file. | | Authority | User and project instructions outrank retrieved material. Evidence never grants authority. | | Evidence | Record typed capability observations, `missing_fact`, `edit_source`, authorization disposition, ordered invocation history, persisted paths/digests, and inspection result. | | Tools | Use the runtime's exposed image tool for image requests (ChatGPT's native tool, or an agent or MCP image tool); otherwise use only observed action-specific capabilities. Never switch on your own to an API-key script, SDK runner, or paid endpoint; name that option and use it only when the user asks. Do not assume network, credentials, or publishing. | | Loop | No open-ended iteration: one immutable brief per image, at most two invocations, and a retry only after the first objective invocation failure. A revision the user asks for after delivery is a new request that changes one thing and restates every invariant. | | Output | One artifact per request: a ChatGPT-native returned image, or a persisted image or prompt-only `.txt` under `.hyper/imagegen//`, with relative path, digest, and inspection status. | | Verification | Capability observations, the ordered attempt history, the persisted path/digest, and inspection of the returned artifact before any visual, rendered-text, constraint, or series claim. | | Stop condition | End at `generated_verified`, `generated_caveated`, `prompt_saved`, `awaiting_input`, or `blocked`; report evidence and limits without fabricating success. | Positive requests: **Korean creation:** “비 오는 밤, 한글 ‘밤 산책’을 선명하게 넣은 서점 포스터를 만들어 줘.” Preserve `밤 산책` exactly once, observe new-image completion, and deliver with a ChatGPT native returned artifact or persisted and appropriate inspection evidence. **ChatGPT creation:** “GPT로 여름 바다 포스터 이미지를 만들어 줘.” Compile the complete prompt, invoke ChatGPT's exposed native image tool once — specifying `gpt-image-2.6-sunburst` when a model selector is exposed, otherwise using the runtime default — and deliver the returned image; do not stop after writing or displaying the prompt. **Authorized edit:** “Use my supplied product photo, keep the label unchanged, and replace only the background with a dawn market.” Confirm authority only because an edit source was supplied, preserve the source, and save a new artifact using observed edit capability. **Fallback refusal:** “Make an illustration, but do not save a prompt if image completion is unavailable.” Record `explicitly_rejected`; unavailable/unknown completion or two failed attempts ends `blocked`, with no prompt write. **Reference-guided creation:** “Make a summer café poster in the style of these two photos.” The photos are style references, not edit targets: label them `Image 1: style reference` and `Image 2: style reference`, generate a new image, and ask no edit-authorization question. Negative requests: - “이 사진 구도를 분석해 줘.” — analysis, not one image or one prompt. - “우리 브랜드용 프롬프트 팩 20개 만들어 줘.” — a reusable prompt pack. - “우리 아이콘 SVG 세트에 맞춰 설정 아이콘 하나 더 그려 줘.” — a code-native vector asset, not a generated bitmap. Boundary requests: - “이미지는 만들지 말고 프롬프트만 파일로 저장해 줘.” — the prompt-only route of this skill, not a different skill. - “포스터를 만들고 그 결과를 평가해 줘.” — create here; evaluating the result is out of scope. | Situation | Route | |---|---| | Every required fact for the brief is present | Compile one immutable brief and act on it | | `subject`, `reference`, `rendered_text`, or `preserve_change` is missing | Ask that rule's one Korean question, end `awaiting_input`, compile nothing | | A supplied edit source | Ask the one authorization question before using it | | ChatGPT exposes a native image tool | Compile and invoke in the same run; name `gpt-image-2.6-sunburst` only when a model selector is exposed | | The user refuses prompt fallback | Set `explicitly_rejected`; unavailable completion ends `blocked` | | No native image tool and no secure write | Put the brief in the response body and end `blocked` | | Supplied images serve only as style, composition, or subject references | Generate with each image labeled by index and role; no authorization question, since nothing is edited | | The user names a project destination ("public/hero.png") | Deliver as usual, then copy the final image there as a new file; an existing file gets a sibling versioned name such as `hero-v2.png` | - Core (`SKILL.md`): routing, contract, trigger routing, workflow, required and forbidden behavior. - Rules: `rules/capability-and-output.md` owns the resolver order, terminals, attempt history, and safe output handling; `rules/prompt-compilation.md` owns brief structure, rendered text, series invariants, and edit boundaries. - References: `references/prompt-crafting.md` (request categories, format selection, text rendering), `references/visual-direction.md` (composition, materials, lighting, series consistency), `references/runtime-capability-and-drift.md` (capability observation and provider drift). - Assets: `assets/evals/hyper-imagegen-cases.jsonl`, the resolver and behavior corpus. - Scripts: `scripts/validate-hyper-imagegen.sh` re-runs the corpus, the resolver self-tests, and the package-document audit. No open-ended iteration. One immutable brief per image, at most two invocations, a retry only after the first objective invocation failure, no reinvocation after retrieval or persistence failure, and no second question once a route has asked its one question. When the invocation cap is reached, reapply `fallback_policy`, and report `blocked` with the attempt history only when fallback is refused or no file can be written. Write the report in Korean unless the user asked otherwise. Keep rendered text, the compiled prompt, paths, digests, and model names exactly as produced, and quote a bilingual report's shared evidence identically in both languages. - Read [rules/capability-and-output.md](rules/capability-and-output.md) before choosing a route and before writing any report or artifact. - Read [rules/prompt-compilation.md](rules/prompt-compilation.md) before compiling the brief. - Read [references/runtime-capability-and-drift.md](references/runtime-capability-and-drift.md) before every image action and after runtime drift. - Read [references/prompt-crafting.md](references/prompt-crafting.md) for category routing, format selection, and the text-render protocol. - Read [references/visual-direction.md](references/visual-direction.md) when visual direction, composition, materials, lighting, or series consistency needs decisions. 1. This core, to confirm the job is one image or one prompt and to pick the route. 2. `rules/capability-and-output.md` for the resolver order, terminals, and artifact rules. 3. `rules/prompt-compilation.md` before compiling the brief. 4. `references/runtime-capability-and-drift.md` before the first action. 5. `references/prompt-crafting.md` for the selected category. 6. `references/visual-direction.md` when direction decisions remain open. 1. **Route, then clarify and compile.** Read the request's signal words, decide `generate` or `edit` (references alone mean `generate`), and pick one generation category or one edit category in `references/prompt-crafting.md`; load only that row. Then apply `missing_fact` before compiling. Ask only the rule's one exact Korean question for a genuinely required `subject`, `reference`, `rendered_text`, or `preserve_change`; do not ask for optional style or model selection. Compile one complete immutable prompt for every resolved image, following the specificity policy in `rules/prompt-compilation.md` (normalize a detailed request, augment a generic one only with allowed details) and preserving exact rendered text, series invariants, and edit boundaries. 2. **Observe and decide.** Load the runtime reference before action selection, then apply the capability rule. An exposed image tool (ChatGPT's native tool, or an agent or MCP image tool) is direct evidence for the generate or edit action and the return channel it declares; route the compiled prompt to it immediately. Ask authorization only for a supplied edit source. Generation never substitutes for editing. 3. **Deliver with bounds.** Invoke at most twice, and retry only after the first objective invocation failure. A successful native return is delivered directly and is not blocked by unavailable repository writing. When local persistence is available and requested, save additionally through the observed secure-write capability using suffixes before extensions (`image-2.png`); a user-named project destination receives a new copy, never an overwrite. File delivery without a native return retains the persistence gates in the capability rule. 4. **Report and stop.** Korean-default reports include terminal, exact invocation count, ordered attempt history, the final compiled prompt, and either native returned-artifact evidence or persisted relative paths/digests. Generated terminals also include inspection status when inspection is available. Never report prompt-only completion for a successful native image return. | Category | Rule | |------|------| | Observed capability | Choose an action from observed evidence, not product names; `unknown` is conservative evidence, never permission. | | Exact text | Preserve rendered text, series invariants, and supplied edit boundaries exactly as the user stated them. | | Evidence record | Keep the typed capability observations, ordered attempt history, persisted path and digest, and inspection result that the terminal requires. | | One artifact | Deliver one artifact per request, written only under `.hyper/imagegen//` (plus a new copy at a destination the user named), and never open an edit source for writing. | | Labeled inputs | Label every supplied image by index and role (edit target, style reference, or insert); only an edit target makes the request an edit. | | Honest terminal | Report exactly one of `generated_verified`, `generated_caveated`, `prompt_saved`, `awaiting_input`, `blocked`, with its evidence fields. | | One question | Ask only the single question the selected route defines; never ask about optional style, format, or model choice. | | Category | Avoid | |------|------| | Prompt as image | presenting a compiled prompt, preview, or plan as a delivered image | | Model improvisation | naming a model the runtime did not expose, or putting a model name in the prompt string or brief | | Extra attempts | a third invocation, or a reinvocation after retrieval or persistence failure | | Unsafe write | writing outside `.hyper/imagegen//` or a destination the user named, following an unverified link, or overwriting the supplied source or any existing file | | Silent escalation | switching to an API-key script, SDK runner, or paid endpoint without the user asking for it | | Code-native substitution | generating a bitmap for an asset that should extend an existing SVG, vector, or code-native set, or drawing an SVG or HTML placeholder when a raster image was asked for | | Unverified claim | any visual, rendered-text, exact-constraint, or series claim without inspection evidence | | Fabricated success | claiming an image or prompt file exists when neither came back this run | - [ ] The request routed to exactly one of create, edit, or prompt-only, and a supplied edit source was authorized before use. - [ ] A missing required fact produced one Korean question and `awaiting_input`, with no compile, invoke, or write. - [ ] The compiled brief came from one category row, followed the specificity policy, labeled every input image's role, and preserved exact rendered text, series invariants, and edit boundaries. - [ ] The action was chosen from observed capability, and a ChatGPT native return was delivered without demanding repository writing. - [ ] Invocation count and ordered attempt history match what actually ran, at most two invocations. - [ ] The terminal is one of the five, with the persisted relative path and digest or the native returned-artifact record it requires. - [ ] No visual, rendered-text, constraint, or series claim appears without inspection evidence, and caveats name what stayed unverified. - [ ] `bash scripts/validate-hyper-imagegen.sh --root .` passes from this skill's folder, or `--root ` passes from anywhere, and `assets/evals/hyper-imagegen-cases.jsonl` still covers the routes this file states. ## Sources > Checked 2026-11-07. | Claim | Source | |---|---| | The rename from `image-maker` and the request to optimize it against public image skills | user instruction, 2026-11-07 | | The code-native boundary, the generate-versus-edit split where references alone mean generate, the indexed input-image roles, the specificity policy, the non-overwriting `-v2` destination copy, and never switching to an API-key path unprompted | OpenAI `imagegen` system skill, (commit `e6afb0d`, file `.system/imagegen/SKILL.md`) | | The resolver, the five terminals, the two-invocation cap, and the safe-write rules | this package's own rules in `rules/`, carried over from `image-maker` |