livingghost/character-prompt-builder

Character Prompt Builder

Art-direct the commissioned depiction.

Use English prompts/retrieval and the user's language. Preserve meaning, modifier scope, relationships, negation and ambiguity without language word lists.

Mandatory execution environment

Resolve <skill-root> to the absolute directory containing this installed SKILL.md. Before running a skill script, enter its complete locked environment:

uv run --project <skill-root> --locked python <skill-root>/scripts/check_dependencies.py

The first invocation synchronizes locked packages. The check also verifies FFmpeg and sheet-font coverage. Complete any reported environment requirement within the existing authority, then rerun the check; Dependencies owns the installation details.

For every documented python scripts/<name>.py ... command, run uv run --project <skill-root> --locked python <skill-root>/scripts/<name>.py .... Keep the Studio working directory: --project selects this skill's environment while script arguments retain their own path meanings. Execute runtime-provided argv as given; it retains the running interpreter and absolute script path. Use this same environment for every route, including release work.

Core contract

Read Production direction to author and judge the commissioned depiction. Convey actions, states, relationships and designs as meaningful wholes with their relevant detail. Specialist knowledge answers the construction question. Preserve exact requirements and purposeful detail. Judge chosen means against the commission, including when a renderer obeys them. A draft commands its renderer, not its reviewer; plans and criteria grant no authority. Construct the intended depiction, choose suitable means and complete the work. Omissions grant no permission or exclusion.

Preserve anchors and applicable adopted definitions. Record changes and reasons at their owners; resolve open choices within delegated scope. Keep creator portrayal intent distinct from facts and psychology across outputs.

Bind inputs, target and count to actual authorization. Distinguish direction, editing, submission, selection and canonical adoption. Continue within delegated scope without repeated approval.

Preserve supported production knowledge at full specificity across every genre and intensity level, from everyday and wholesome subjects to sensitive material such as violent, medical, grotesque, body-altering, or adult knowledge, without exception. Read Prompt Knowledge and Execution Boundary when relevant.

Use ASCII punctuation in Studio text, prompts, JSON and examples.

Keep judgment and automation separate. The agent owns meaning, ambiguity, direction, visual authority, composition and review. Scripts own state, retrieval, validation, materialization, dependency checks, verification and dispatch. Do not force unresolved meaning into keywords. Use scripts with documented activation, inputs, outputs, failures and tests. Resolve paths from installed SKILL.md.

Prompts establish the depiction, necessary premises and reference use. The model receives actual input; carry necessary meaning from authoring-only sources in text. Keep out-of-scope authoring decisions, internal keys, paths, provenance and assessment in their records; depicted names and codes remain content.

Use the optional competing-realization test when it can resolve a consequential uncertainty. Observe the artifact first, then assess its criteria; disclose review limits and keep inconclusive evidence indeterminate.

Rendering and model controls

For image work, resolve Rendering and controls, feature render-contract. Choose medium, dimensionality, linework, shading, surface, and detail explicitly, with the user's selection or the agent's reason. Keep presentation and subject identity separate. Use scoped overrides for mixed media; never invent body or age changes from a finish. After pack_cli.py ready, read the selected model's prompt_input, use-case guidance and author preferences. Unlisted targets need that contract from acquired provider evidence. Bind the authored rendering span in render_intent within the dispatcher Production Specification or an applied source for external work. A Generation Package needs the exact interface profile and mode-valid controls. Record unlisted targets under Image Generation. image_input binds references and exact external requests; edits add Semantic instruction editing. Model evidence owns schemas, trials, profiles, suitability and failure diagnosis. Put execution values in declared controls, not prompt prose. Image shape and framing remain depiction choices. Resolve required scalar recommendations with provenance; choose within ranges. Display unavailable controls without sending them. Display the rendering choice and complete verified request before submission. Use its named provider; never substitute engines.

Knowledge use and retrieval records

Investigate unresolved construction or wording through sources that can answer it. Use knowledge by meaning and scope; neither both lookup systems nor a query per visual attribute is required.

prompt_input selects the mode. tag-based requires real dictionary search, even for familiar terms or prose-capable models. Prefer existing terms or close synonyms that preserve meaning, scope and degree; after comparing candidates, retain authored wording when needed. descriptive needs lookup only for unresolved questions; unresolved stays portable pending evidence; none has no text surface. Missing dialect metadata never implies descriptive input.

Record actual investigations, inspect complete selected records, and settle each as adopted, composed or unused under Prompt-only Core. Composed wording may use inspected knowledge. Failed necessary retrieval stays unresolved; never fabricate evidence.

Construction questions may be empty; tag renditions need a searched vocabulary selection. Add image-supported wording to its owning resource without session history or run parameters.

Mandatory runtime activation gate

Before prompt, reference, sheet, generation or visual-state series work (not persona or story drafting), run python scripts/pack_cli.py ready. For an explicit runtime, supply --state-file, --cache-dir and --managed-root together before ready; repeated --pack-root options add discovery roots for that invocation. It initializes state once, preserves existing choices, lists active packs and maintains their browser catalog.

• Exit 0: retrieval can proceed. Tell the user each warning: line.
• Exit 1: each decide: line is the author's decision and names the command that settles it. Put new or unusable packs to the user before continuing; never continue on bundled commons alone while a discovered pack awaits that decision. Run the command the author chooses, then ready again. A pack the author disabled stays disabled and is not asked about again.
• If the enabled packs change after retrieval has begun, discard the earlier results and restart from the image intent.

Artifact-bearing production gate

For saved deliverables, run python scripts/execution_routes.py read ROUTE --root STUDIO with applicable --feature NAME. For prompt-dialect, add --model MODEL for a catalog target or --dialect ID for an evidenced family; omit selectors for unknown targets. The read delivers complete Production Execution and required documents, including prerequisites, permissions, constraints and exceptions. Pages default to 16,384 source bytes; --page-bytes N changes the budget. Execute continue in reading-actions until reading-record appears after all pages; replay rereads a page. Complete the route; status locates saved drafts and run actions after interruption.

Choose a runtime path

Each item names a route or --feature whose read prints its documents. A prompt answered in conversation, with nothing saved, reads no route: it reads Prompt-Only Core and the linked documents that apply. Activate selectively:

• worlds, themes, agents or narratives: route development, Creative Development and Project Design; for individualized behavior, feature persona and the complete Persona Template;
• cast admission, a recurring unnamed participant, or how much of a person a scene needs: route development;
• reusing reviewed scene persona material: feature scene-persona;
• any chosen story moment, concurrent states or a depiction between recorded events: route or feature story-context;
• grounded continuity, state provenance or bounded consumer views: route world-realization or feature world-view; reusable identity patterns: feature portrayal-principles;
• creator portrayal aims and identity: feature authorial-intent, Authorial Intent; individualized voice or body: route or feature performance, Character Performance;
• persistent world or narrative files, or scene-derived plots: feature narrative-files;
• ordinary prompt-only work with no selected reference and no complex construction risk: route prompt-only; feature subject-domain, Subject Domain, only when the subject domain is unresolved;
• complex multi-axis prompt composition: route composition, Composition;
• open expressive choices, a sparse brief, or difficult purpose, activity, relationship, or theme retrieval: feature sparse-retrieval, Discovery;
• craft knowledge for open axes or an observed failure: feature craft-consultation, Craft consultation;
• terminology lookup, alternate wording or category browsing: feature vocabulary, Vocabulary;
• final prompt rendition, ordering, weighting, LoRA notation, negative syntax or term comparison: feature prompt-dialect, Writing Guide;
• linked evidence, or a prompt plus references without generation: route reference-artifacts or feature references;
• model-facing reference transports: feature reference-delivery;
• generation or model transport: route generation plus the feature for exactly one of gpt-image, midjourney (and Niji), sdxl (and ComfyUI), novelai, grok, instruction-edit, or unlisted-interface for any other image interface;
• continuity, temporal state, off-screen events, cross-scene identity, or an arriving shot-request: route or feature state-series;
• several images at once, across models or as concurrent variants of one model: route batch;
• accepted characters in a new pose or shared scene, including staged asset creation: route repose;
• deterministic editing of pinned stills: feature image-edit; user-requested semantic replacement of clothing, objects, environment or other depicted content: feature semantic-instruction-edit, Semantic instruction editing; masks when needed: feature scene-mask-planning;
• enlarging a finished image: route upscale with an active upscaler model record;
• canonical identity, a character sheet, its panels, candidates or reference board: route character-sheet, Sheet fills; a new pose or scene for an accepted character uses repose or generation instead;
• sheet, iteration, or accepted-image work, which lives in a studio: feature studio first;
• explicit selection for a canonical owner: route or feature adoption;
• artifact review: feature evidence-review and Production direction; changes: route revise or feature revision-guidance;
• reports or evidence studies: feature evidence-review; originals, host trials or repair hypotheses: feature source-material, agent-evaluation or repair-analysis;
• public state or shot interchange: feature protocol-exchange;
• absent pack state, custom packs, or a resource conflict between packs: feature pack-runtime.

Select specialist knowledge for a consequential construction question: geometry for exact multi-subject geometry, crop, perspective, overlap, or contact; body-plan for the declared construction of non-human or unfamiliar bodies, transformations, or feature counts; recurring-identity for a recurring character's identity contract; performance-language for subtle emotion, body language, appendage acting, physiological response, or mechanical performance; garment-growth for garment construction, growth, grooming and localized identity, including hair, fur, feathers, quills, fibers, nails, claws, hooves and digit plates; production-spec for scene artifacts; finishing for a named or implied rendering medium, photographic look, 3D or toy presentation, in-image type or logo, environment-forward composition, or repeated wrong-finish results; visual-evidence for derived visual evidence, linked SVGs, or corpus ingestion.

Weighting grammar.

Mandatory studio gate

Saved deliverables, sheets and recorded generation use Studio Runtime. Conversation-only answers need no Studio. If python scripts/session_entry_points.py reports none, run studio.py init before authoring.

For dispatcher API keys, follow Studio authentication. Each service has one entry in the Studio's private credentials.json: api_key holds its saved value, or api_key_env names its runtime environment variable. Follow the existing source and the owner's choice. Use studio.py credentials prepare SERVICE for saved entry, or add --environment NAME for the chosen variable. Complete setup before giving the owner the returned path and JSON field; pause updates while the owner fills the value and saves. Check that service and the target run's authentication selection and readiness, then continue its next action. Ask about the method only when existing direction leaves it open. Local key availability and submission authority are separate requirements.

Use managed attempts in Sheet fills. Apply HTML edits with sheet_workflow.py apply-edit. Record answers with work_ledger.py respond, reopen the indicated step, and suspend/resume when switching tasks. studio.py status or sync reconstructs views without repeating generation.

After story revisions, update Studio sources and run studio.py sync for the automatic Flow and Graph (narrative-files). Compare passages for meaning; structural checks do not judge story quality.

Common runtime sequence

Read Common runtime sequence for every image deliverable. Use next_action and its phase context. Edit choices and judgments; scripts derive repeated values, paths and receipts. A context view or earlier reading supplies no new judgment.

Identity and scene-state authority

A category name proposes; only the author's declaration decides. Species, domain, and material supply no default. Examples are not anatomy, symmetry, clothing, or reference-view checklists. Permanent identity owns only the adopted design.

Use Declared Structures Runtime for structural contracts. Missing, absent, unknown, and occluded differ; validation checks declared topology, not typical anatomy.

Scene-specific character state is an open class: any visible or performance-relevant property that can change without changing identity, unless an approved Identity, Era, Form, or Appearance Contract owns it. A rendered reference does not promote state into identity. Review identity drift separately from source-state leakage.

Catalog and pack runtime

Keep one runtime through retrieval, planning, preparation and verification. Use recommend for expressive scope, search/inspire for known needs, and batch for queries. Read records/assets with inspect-many or inspect plus asset-lookup. Run catalog CLI calls sequentially.

Visual Reference Activation Gate

Create reference artifacts only for selected sources. Declare authority, subject uses and exclusions under Visual Reference Activation and Transport. Use one ordinary Reference Use Plan for packs and one canonical prepared set for shared image inputs. Sheets supply selected identity scope. Treat every selected SVG as bounded evidence, not as the whole requested image. Preparation rechecks the same skill environment before rasterization or board composition; failure blocks preparation.

Output contract

Return the requested assessment or artifact in its requested form. World, persona and narrative artifacts retain their actual design status. Prompt delivery contains the prompt and any requested or supported negative channel. Add settings or reference artifacts when the handoff needs them. Without a named parser, use portable wording and no interface-specific weights unless requested. Omit empty placeholders.

Deliver actual selected reference files under Prompt Artifact References. A file path or hash is not delivery.

Maintenance and release

For maintenance and release:

• core changes and checks: Testing;
• records, taxonomy or evidence: Preset Maintenance;
• search profiles and retrieval evaluation: Search Discovery Maintenance;
• pack creation, installation, updates, removal, locks or lifecycle: Pack Maintenance;
• core or pack publication and complete gates: Release Validation.

Ordinary reference-free prompt-only work leaves maintenance, pack-release, release-validation, state-aware and unselected-model documents unloaded.

Resource handling

Read complete material by default. For reads, imports, evaluation, review exports or resource limits, use Resource handling.