Prompt compilation

Before final delivery of a substantive authored, revised or generated creative artifact, automatically apply output review. Reuse domain QA in one bounded loop; inspect actual media, preserve accepted locks and stop unchanged on a pass. This does not run for greetings, menus or onboarding questions. When Superpowers is in the skill list, load superpowers:systematic-debugging before diagnosing a reproducible technical defect and superpowers:verification-before-completion before calling a deliverable ready; Studio keeps ownership.

Out of scope first. If the request belongs to another owner, load that owner's skill and continue there; do not produce its deliverable here. A complete poster, flyer, ad or other designed graphic with copy, price or logo goes to Graphic Design. With essentials missing, ask one question (the subject); default the rest and label the defaults.

Read the compiler. For the prose itself, follow narrative image prompts (checked 2026-10-25): the order of clauses, the length per model, exact text, edits, exclusions and which tools rewrite prompts. Apply the conditional public research gate before committing to a substantive prompt. Search for visual or contextual inspiration only when the user asks for it or the subject depends on real-world facts. For a named model, research is triggered: load mapping and evidence, actively check current owner guidance and first-hand practitioner evidence for the exact family/version/surface, and state any gap. No generator name means a conservative natural-language prompt, not an invented default external provider.

For gpt-image-2, the prompt examples hold 38 licensed, credited worked prompts (portraits, social posts, game UI, infographics, posters). Consult one or two for the model's structure after the direction is chosen; write a new prompt from the brief and keep an example's credit line if any of its text is reused.

Read visual subject craft when the requested image materially depends on human plausibility or identity, wardrobe/accessory construction, real-product truth or claims, photographic camera/light/material coherence, camera reframing, or a still image intended as a possible image-to-video starting frame. Retrieve only the sections relevant to that task; do not route unrelated abstract or purely typographic work through it. For general stills, use only the visual-craft decisions that matter; do not force poster copy, grid, or eight-priority static-graphic stages where there is no communication layout. For finished text-bearing static graphics, retain the integrated one-pass and exact-copy contracts. The reference governs still-image craft only and does not replace source binding, current generator research, or provider authorization.

For a transfer combining identity, garment, product or pose sources, use the applied reference-transfer playbook and its worked source-bound examples. For a photographic shot whose focus, contact, lighting or surface response needs concrete construction, use capture and material recipes; the optional still shot card records only decisions needed by that frame. These are conditional applications of the existing craft, not additional stages or render calls.

When the user gives an image and asks for a prompt that recreates it (often with their own face, character, animal or product as REF), follow reference reconstruction: the model image is analysed and never becomes a REF, the mechanism comes first (an optical effect by its physical cause: a mirror shows a reflection, not a framed picture), space is described as measured or stated relations with each figure's gaze and focus, colours have roles, the image probe runs whenever a shell or code execution is available, and a template the user supplies is filled exactly.

For recurring photoreal humans or identity-critical characters, use the human reference capture card. Ask for one missing view at a time and only when it affects the planned angle, expression or wardrobe; bind each image to the properties it actually shows. Preserve ordinary visible traits instead of beautifying the person. When choosing a model, research the exact current generator surface and operation rather than assuming every image model handles realistic faces or multi-view identity equally well. When a real-world reference would sharpen the prompt and the user has none, take one from the Open Assets reference base as a reference only (never a real person's identity), with its licence.

Extract an adequate direction from a complete user brief. A formal upstream file is not mandatory. If creative choice or assistant-drafted copy remains unselected, return the narrow missing choice, not a supposedly final prompt containing alternatives.

Keep conversation, technical prompt and visible-copy languages separate. Default technical prompt to English; follow an explicit language instruction. Quote final copy verbatim. Compilation is not a copywriting or translation pass: preserve selected text from Humanizer, screenplay or lyrics workflows exactly wherever the user approved it for visible use, and do not turn unrequested dialogue/lyrics or handoff metadata into artwork text. Explicitly approved line breaks are locks; otherwise describe permitted reflow without changing words or punctuation. Never add decorative dates, pseudo-logos, partner marks, arbitrary microcopy or imagined facts.

For a finished text-bearing raster that stays here (text inside a photographic or illustrative still, or a graphic Graphic Design hands over for compilation) compile one integrated instruction for the entire graphic, including text, unless the exact-copy route is chosen and named: then the prompt is for a background without text that reserves the text zones, and the static compositor sets the copy. The eight construction priorities are semantic organization, not eight renders, layers or mandatory long headings. Do not quietly replace it with text-free art, SVG, HTML or a later overlay. Explicitly requested layers, vector or code need a separately agreed output contract.

The actual generation tool owns callable parameters and model selection. Semantic requests are not API fields. Never import flags from a different family or wrap prose in JSON and call it official syntax. When the user asks for a JSON prompt, the target model has its own schema (Ideogram 4.x, Bria FIBO) or one record must feed several prompts and edits, follow structured image prompts: one Studio planning record, compiled into the model's own JSON where one exists and into ordered prose elsewhere, with `prompt_record.py` to check and compile it where code runs. Do not ask about API/wrapper unless the user raises it.

When the request names unwanted microtexture or a concrete artifact-prevention problem, read periodic microtexture and preservation. Apply only relevant material requirements in the complete prompt; do not add a universal suffix or claim an unseen render was inspected. Before returning it, remove blacklist items for unreported patterns: target the reported defect, not all geometry or repeated detail.

For a cut-out asset that a motion film will animate, follow generated cut-out assets: one object, straight-on, moving parts clean, no text, and the background its generator needs (a transparency parameter where the tool has one; for Gemini, never "transparent background" but white then an edit to black, or flat green for a solid subject; light effects on black with no backing).

Before release

Check exact text inventory, source binding, preserved concept, plausible proportions, attention order, useful detail, compatible instructions, method and next review test. Resolve a failed critical gate instead of merely adding “professional quality” or a long negative list.

Return the complete prompt in one fenced block. Unless the user requests prompt-only, add a short production note with format, first read, protected assets and checks. Stop there for prompt-only. If the workflow still calls for generation and it has not been requested, a concise consent question is appropriate.

When generation is requested, invoke only the real authorized host tool, following its own display and execution rules. If unavailable or errored, report that state accurately and preserve the usable prompt. Do not invent a policy refusal, silently retry or select a paid fallback. Return with the actual image for output review.

Integrated workflow contracts

Use the integrated workflow-kit method for this owner’s artifact fields, bounded review and downstream handoffs. Read Studio integration authority for precedence and the single active route. Preserve this owner’s domain craft and the user’s requested stage.

Revisioned prompt handoff

For strict locks, several requests, target adaptation or execution readiness, use Creative Prompting Standard and its contract template. Bind each request to its actual attachments, exact copy, source revisions and required continuity. Keep requirement, readiness and inspected outcome separate; a planned end frame cannot be an accepted carrier. Final prompts stay complete and standalone, each in its own fenced block.

Identity evidence and capability mapping

For recurring people, preserve the identity workbook decisions and use reference capability routing for the chosen model/surface. For a board request, compile the production sheet contract without inventing source frames.