Limner

A framework for taste-enabled creator agents — three instantiations across pixel art, ASCII rendering, and hardware cartridges. Agents that judge their own work and refuse to ship below the bar.

Taste-Enabled Creator Agents·Universal Persona Systems
Taste-Enabled Creator Agent Framework Active — Three Instantiations Autonomous Agents Quality Gating Multi-Tool Orchestration Computer Vision
Limner image asset pipeline

Published:

Context

Most creative AI tools produce output and stop. The human judges whether the output meets a bar. That's a dumb division of labor — it means every batch requires human review at commodity rates, and the AI never learns what "good" means in any durable sense. Limner inverts the pattern: the agent carries the taste criteria, rejects its own output when it fails those criteria, and only surfaces work that meets the bar.

Approach

Limner is a framework, not a single agent. The framework provides the core machinery — taste-evaluation loops, self-rejection gates, multi-tool orchestration, quality-gated output pipelines — and each medium-specific instantiation supplies its own taste criteria, tool chain, and rendering conventions.

Three instantiations currently exist or are in development:

  • Limner: Pixel — pixel art asset production. Active production; multiple batches of 500+ uniformly stylized assets delivered autonomously.
  • Limner: ASCII — textmode rendering (BBS-era through modern ASCII UI). In development.
  • Limner: UPS Cartridge — a hardware-substrate variant. Console-vs-Cartridge architecture on commodity hardware. v1 shipped; v2 in development.

The instantiations share the core framework. They differ in medium, taste criteria, and the substrate they run on.

Limner: Pixel

Status: Production

Summoning Chamber needed VGA/pixel-art visual assets in a consistent style across hundreds of pieces. Commissioning human artists at that volume was cost-prohibitive. Running MidJourney and hand-culling outputs was volume-prohibitive. I needed an agent that could produce uniformly stylized assets and reject its own non-conforming output before delivery.

Multi-stage production pipeline with autonomous quality gating:

  1. MidJourney for concept exploration
  2. PixelLab (via custom Limner agent) for pixel art conversion
  3. Automated quality gate — agent evaluates output against stylistic criteria and rejects non-conforming results before they reach the output queue
  4. Vectorization for print-ready final output

The quality gate is the critical piece. Without it, this is just a chained tool pipeline. With it, the agent is making aesthetic judgments at each stage and refusing to pass through work that fails those judgments.

Multiple batches of 500+ uniformly stylized assets produced with zero human intervention required. Successfully applied to game assets, UI elements, icon sets, and vanity portraiture. First production-scale exemplar of the Limner framework.

The lesson: the value isn't in the generation step, it's in the rejection step. An agent that will throw away its own work is a different creature than one that will ship anything.

Limner: ASCII

Status: In Development

ASCII rendering online is dominated by commodity converters that treat glyphs as darkness-values and produce output that's technically ASCII but aesthetically indifferent to the medium. The textmode tradition treats characters as characters: shapes with their own geometry, weight, and inherent visual logic. A taste-enabled ASCII agent needs to know the difference.

Limner: ASCII extends the framework with ASCII-specific taste criteria:

  • Glyph-geometry awareness — characters as shapes, not darkness-values
  • Negative space discipline — the whitespace does half the work
  • Character-set consistency within a single rendering
  • Fidelity to rendering traditions — classic ASCII, BBS-era, box-drawing, textmode/ANSI art

Four production domains: classic BBS-era ASCII studies, image-to-ASCII conversion with taste gates, ASCII patterns and animations, ASCII UI components.

Deliverables in progress: Astro UI kit with ASCII-rendered components, editorial image series, style guide, reusable AI production skill.

The test of the Limner framework is whether its core machinery transfers cleanly to a second medium. If it does, the framework is real.

Limner: UPS Cartridge

Status: v1 Shipped — v2 In Development

The Universal Persona System treats portable agent identity as a self-contained artifact — a Console-vs-Cartridge architecture where a stateless runtime reads an identity package to boot a specific agent. Limner: UPS Cartridge applies the UPS spec to the Limner framework and packages it as shippable hardware.

v1 (shipped): A live Limner agent runs on a USB-C SSD housed in an ASUS Cobble enclosure — IP55 aluminum enclosure, dual M.2 interface, tool-free installation, genuinely pocket-sized. Plug it into any host machine, and the Limner persona boots from the drive. The Cobble's cobblestone form factor makes the Cartridge literal: a warm aluminum stone that carries a hand-raised agent.

v2 (in development): Dedicated compute + I/O via Raspberry Pi 5 housed in an Argon ONE V5 Dual M.2 NVMe case. Dual NVMe slots, cast aluminum, PWM active cooling, DAC, dual HDMI, HAT compatibility, ZigBee support.

v2 hardware → UPS mapping:

  • Dual M.2 NVMe slots → Console/Cartridge split — runtime on one drive, writeable state on the other
  • Pi 5 compute → Console layer — local inference or cloud relay with persistent local state
  • DAC + 3.5mm audio → Agent can hear and speak
  • Dual HDMI + HAT + ZigBee → Agent body — display, sensing, actuation
  • Cast aluminum enclosure → Discrete physical artifact, holdable, pocket-pet scale

Target for v2: a Limner instance running on dedicated hardware, Console and Cartridge on separate swappable drives. Swap the Cartridge drive — same Console boots a different agent.

The cartridge stops being a metaphor.

Commercially-viable digital pocket pets.

Outcome

Limner originated as a sub-project spawned by Summoning Chamber when that project needed VGA-style visual assets. What was supposed to be a tool became the thesis: agents that carry aesthetic judgment are categorically different from agents that execute tasks. The framework is the first generalized expression of that thesis, and each new instantiation tests whether the core taste-evaluation machinery holds across mediums and substrates.

Taste is not a genre. It's a discipline.