An open-source information architecture for commercial visual-asset workflows: it turns brand style guides, creative rules, and workflow constraints into machine-usable normative structure for AI-native visual asset production.
This is not a better image-generation interface. It is a structured definition layer for making visual-production intent, constraints, references, outputs, and governance machine-usable.
Image models provide latent image-making capacity; this repo structures the production ask around that capacity — what image is needed, why, under what constraints, with which references in which roles, and how the result becomes governable.
The reusable object is a portable asset-definition package — the definition layer for a commercial visual-asset workflow, designed to be substrate-independent: product truth, business and creative intent, constraints, required outputs, slot variation, visual references, decision ownership, approval boundaries, capture/governance rules, and reuse/family-coherence rules, carried as each workflow needs them. The repository is a current worked reference architecture — a prototype proof of that layer, refined as evidence accumulates, not a finished product spec.
The wedge is whether style guides, approved references, and creative rules can be made legible as normative information structure that a repo-local pipeline can carry and verify without collapsing into vague creative direction.
Product and SKU data integration is expected, but it is not the wedge. The wedge is the normative structure that makes creative intent, constraints, references, outputs, and governance usable above it.
The repo is building a durable definition layer above changing tools and models, not a generation interface or runtime orchestration layer. It articulates a layered information architecture across brand-system, category, mode-specific, packet, slot, candidate-generation, curation-seam, and governed-asset layers, and pressures that architecture against four worked-example workflow modes at uneven evidence depth:
- SKU-driven product imagery — operationally grounded at full-flow depth
- collection / merchandising — operationally grounded at full-flow depth
- brand campaign / editorial — bounded full-flow plus structural-proof depth
- marketing / message-driven — message-hero full-flow proven in-base; support slot held
Still-image first; video extension is downstream of the layered model stabilizing.
The 2c marketing hero now demonstrates the full chain end to end inside the system: the definition layer drives generation, candidates accumulate, a human curator selects one, the selected candidate is captured into a governed asset with no external import, and the output slot is accepted. Across the other worked modes, the same definition-layer structure is pressured at uneven evidence depth. Airtable is the worked substrate that demonstrates this — one substrate, not the deliverable. The deliverable is the portable definition layer above it, designed to be substrate-independent.
Target vs earned. "Machine-usable" and "substrate-independent" name the target, not present-tense generality. Earned to date: one package re-expressed outside Airtable and recovered from a file-bundle second substrate, and one bounded, model-mediated package-instantiation planning consumer calibrated on one known-achievable filesystem target. Not yet earned: operational cross-substrate execution, a reusable consuming grammar, a general converter, a validator, or a runtime. "Machine-usable" here means model/advisor consumption of the definition layer to generate fit-for-purpose artifacts — not a validator or runtime orchestrator.
The strategic reader path — to the concrete object first, then depth:
docs/problem-and-approach.md— the differentiated failure this repo addresses, and the object distinction it makesdocs/asset-definition-package-map-2c-marketing-hero-v1.md— the concrete package: one chain end to end (intent → discretion → variance → selection → governance)docs/package-lifecycle-partition-v1.md— the three lifecycle states a package spans (prospective definition / execution-run state / governed record) and how to read its statusesdocs/architecture.md— current architecture front doordocs/method.md— ontology / application split and working sequence- structured IA + deeper artifacts —
docs/structured-ia-model-v2.md(canonical machine-readable IA; hybrid prose + YAML appendix),docs/layered-reference-and-discretion-architecture-v1.md(nine-fold reference-function taxonomy + nine-site discretion decomposition), anddocs/brand-system-input-cross-category-synthesis-v1.md(Phase 2 six-category cross-category synthesis)
For a first reader, the simplest complete chain example is the 2c marketing hero — a single output slot carried end to end, not the whole marketing mode:
For the cleanest layered IA architecture case, see the SKU-driven mode, anchored against PKT-SKU-009 and PKT-SKU-010:
docs/index.md— full navigation across all repo artifactsdocs/milestones/— plateau markersAGENTS.md— repo-local execution guidancedocs/diagrams/asset-pipeline-ASK_architecture-tree.html— visual architecture tree at orientation depth; regenerated when a plateau milestone marker lands or an equivalent architecture-front-door shift makes the diagram materially stale, so its update cadence is slower than the prose docs; the prose docs are authoritative if the two diverge
Prototype test bases (worked evidence surfaces; Airtable is one substrate, not the deliverable):
Airtable base // Integrated Four-Workflow Home-Goods v1— the base the 2c marketing-hero chain runs in; the worked evidence surface for the end-to-end in-base proof (definition layer → generation → candidate accumulation → selection → governed asset, no external import)Airtable base // SKU-Driven Furniture v1— original prototype base for SKU-driven Furniture v1 and early cross-mode pressure workAirtable base // Campaign-Driven D2C Home-Goods v1— separate campaign-mode prototype base for cross-base / cross-brand pressure work
Background article series, in chronological order:
- From Brand Rules to Creative Discretion
- From Normative Structure to Execution
- The Beautiful Picture is Not the…
- The Creative Act of Selection
- From Execution Proof Back to Normative Structure
- Creative Discretion Is Not Creative Intent
- The Missing Layer Is Not Generation
- Selection Is Not Governance
- The Image Is Not One Thing
- The Model Does Not Know What Image to Make
- A Reference Does Not Carry Its Own Instructions
- Synthesis Is Not Authorization
- The Manual Barrier Was Never the Craft
- The Hollow Middle Is Not Empty
- Provenance Is Not Authorship
Copyright 2026 Andrew S Klug // ASK
Licensed under the Apache License 2.0 // see LICENSE