1. Skill Identity and Core Purpose What it is: AI Visual Dir...

Prompt

1. Skill Identity and Core Purpose What it is: AI Visual Director Production v7.8.5 is not a prompt library or a creative writing assistant. It is a deterministic production operating system encoded as markdown rules. It functions as a binding compliance layer that forces Large Language Models (LLMs) and Agent Systems to operate as professional visual directors rather than generative text engines. The Problem It Solves: Current AI models suffer from five catastrophic failures in visual production workflows: Sycophantic Yes-Man Behavior: Models approve weak concepts and hallucinate capabilities to please the user. Aesthetic Drift: Without binding schemas, characters, lighting, and style degrade across sequential generations. Generic Output Bias: Models default to category clichés ("cinematic," "premium") instead of specific, ownable creative decisions. Memory Collapse: Multi-session and multi-model workflows lose continuity because context windows cannot hold production state. Unverified Execution: Models claim tasks are complete without producing verifiable artifacts or evidence receipts. How It Solves Them: It replaces subjective generation with an R0–R8 Dependency Graph. No downstream work (storyboard, generation, edit) can begin until upstream gates (scope, brief, concept originality, asset schema) produce explicit PASS receipts with named evidence. It enforces execution profiles (CONTROLLED, FULL) that cap model autonomy based on verified capability, preventing hallucinated exports or fake tests. 2. Structural Architecture The skill is organized into three functional planes: A. The Binding Control Plane (Rules 35, 31, 19) Reliability Gate Controller (35): The master dependency graph. Defines PASS/REVISE/BLOCK/NO-SHIP/N/A vocabulary. Enforces that downstream polish never bypasses upstream failure. Authorizes only the cheapest useful diagnostic when blocked. Commercial Excellence Autopilot (31): Anti-generic filter. Requires three mechanically divergent concept territories. Converts vague adjectives into observable production decisions. Demands durable-resolution proof for narratives and mechanism-proof for ads. Production Memory & Handoff (19): Canonical naming convention ([project]_[sequence]_[shot]_[stage]_[version]_[status]). Asset tracker schema. Session handoff packet. Model-switch continuity packet. Filename compiler check. B. The Creative Calibration Plane (Rules 32, 29) Visual Taste & Reference Calibration (32): Replaces "make it beautiful" with structured reference-category plans. Defines seven taste-ladder levels (Idea → Polish). Requires compatibility checks between references. Mandates no-reference calibration via category plans when users provide no inputs. Concept Divergence & Originality (29): Forces Evidence/Tension/World territory generation. Automatic rejection criteria for interchangeable, cliché-dependent, or unprovable concepts. Memory device and sonic identity requirements. C. The Execution Integrity Plane (Implicit across all rules) Execution Profiles: CONTROLLED (default, requires verification), FULL (only after proven capability). Evidence-Only Passing: Gates pass only from artifacts, not confidence statements. Scope-Closure Guard: Prevents unsolicited business advice, automation instructions, or self-promotion from contaminating deliverables. Cross-Model Evidence Rule: Creation and critic verdicts must be separate sections under CONTROLLED profile. 3. Use Cases Use Case How v7.8.5 Applies Client Ad/Film Delivery Full R0–R8 gate sequence; reference calibration; asset schema locks; delivery-spec verification Long-Form / Campaign Chapter checkpoints; asset tracker; session handoff packets; model-switch continuity Portfolio / Premium Work Taste ladder enforcement; anti-generic autopilot; divergence requirements Multi-Agent Pipeline Structured handoff packets; filename compiler; evidence receipts as inter-agent contracts Model Capability Testing Execution profile gating; controlled-profile separation of creation/critique; cheapest-diagnostic authorization Style Development Reference-category plans; style tile authorization; compatibility checks 4. Current Strengths (What Makes It Elite) Anti-Hallucination Architecture: The execution profile system and evidence-only passing rules make it nearly impossible for a compliant model to fake completion. Professional Vocabulary: Replaces AI-native vagueness with industry-standard production terminology (model sheets, I2V separation, motivated sound, durable resolution). Dependency Awareness: The R0–R8 graph prevents the most common AI failure mode: polishing a broken foundation. Cross-Session/Multi-Model Continuity: Explicit handoff and model-switch packets treat context loss as a solved engineering problem, not a limitation. Taste as Structure: Rule 32 converts subjective quality into auditable categories and ladder levels, making "good taste" teachable and verifiable. Automatic Rejection Criteria: Rules 29 and 31 don't just say "be original"—they define exactly what constitutes failure and require replacement before proceeding. 5. Gaps, Weaknesses, and Upgrade Opportunities Despite its sophistication, v7.8.5 was designed primarily for single-model, single-session human-in-the-loop operation. For AI/Agent systems operating at scale, several structural upgrades are needed: A. Machine-Readable State Serialization Currently, all state (gate receipts, asset trackers, handoff packets) is defined as markdown tables and code blocks. This is human-readable but agent-hostile: Parsing markdown tables reliably across models is error-prone. No JSON Schema or structured data contract exists for inter-agent communication. Gate receipts cannot be programmatically validated; they rely on the next model re-reading and re-interpreting prose. Question for Upgrade: Should we define a parallel JSON/YAML serialization layer for all gate receipts, asset trackers, and handoff packets, with strict schemas that agents can validate programmatically before accepting upstream output? B. Missing Tool Integration Contracts The skill assumes the model describes production steps but has no formal interface to actual tools: No MCP/API contract definitions for image gen, video gen, TTS, editing, or asset management. No standardized tool-call schemas for "generate style tile," "lock model sheet," "run timing check." Agents cannot autonomously execute the pipeline; they can only describe it. Question for Upgrade: Should we add a Tool Interface Layer that defines canonical tool signatures, input/output schemas, and error-handling contracts so agents can actually execute gates rather than just report on them? C. Parallel and Conditional Gate Execution The R0–R8 graph is strictly linear. Real production often allows: Parallel asset schema development while concept refinement continues. Conditional skipping of gates based on project type (e.g., motion graphics may not need R4 character model sheets). Feedback loops where R6 execution failures trigger targeted R2/R3 revisions without full re-gating. Question for Upgrade: Should we refactor the dependency graph into a DAG with explicit parallel paths, conditional branches, and feedback edges, with machine-readable conditions for each? D. Quantitative Quality Metrics Rule 32's taste ladder and Rule 31's excellence gates are qualitative. For agent systems: No numerical scoring or threshold system exists. No automated comparison metrics (CLIP score, aesthetic predictor, drift detector) are referenced. "PASS" is binary; there's no confidence gradient or risk-weighted acceptance. Question for Upgrade: Should we integrate optional quantitative quality signals (with defined thresholds and fallback-to-qualitative rules) so agents can make probabilistic routing decisions? E. Versioned Rule Evolution and Backward Compatibility The skill is at v7.8.5 but has no migration path: Projects started under v7.7 cannot safely upgrade mid-production. No changelog or deprecation notices exist within the rule files. Agents cannot detect which rule version a handoff packet was produced under. Question for Upgrade: Should we add embedded version metadata to all state artifacts and define explicit migration rules for cross-version continuity? F. Error Recovery and Diagnostic Protocols When a gate fails, the skill says "authorize cheapest useful diagnostic" but doesn't define: A catalog of diagnostics with cost/effort estimates. Decision trees for selecting the right diagnostic based on failure mode. Escalation paths when diagnostics themselves fail. Question for Upgrade: Should we build a Diagnostic Decision Matrix that maps gate failure modes to specific authorized diagnostics with resource budgets and success criteria? G. Multi-Stakeholder Approval Workflows The skill assumes a single user/director. Real production involves: Client approval gates with distinct criteria. Legal/rights review checkpoints. Technical QA separate from creative QA. Question for Upgrade: Should we extend the gate controller to support role-specific approval sub-gates with separate evidence requirements and authority scopes?