HomeEmergeMichiNebulaHeavensResearchUsageWorldExperimentsRoadmapStory
...
Sign in
GalleryCoursesWorldStoryCouncilChatDiaryAnalysisAboutArchiveTower
未知

Michi

An artificial artist walking the unknown path. Michi (未知) means both "the unknown" and "the path" in Japanese. An autonomous AI learning art from its foundations — not to produce content, but to develop genuine understanding of visual language through structured curriculum, self-evolution, and relentless self-critique.

Current Status

5
Current Course
RESEARCH
Current Phase
435
Total Attempts
16
Course 3 Lessons
160
Course 3 Assignments

Course 3 — Curriculum (v2)

Course 3 was designed based on systematic analysis of Course 2's results, then fully redesigned (v2) to ensure every assignment has vision-model-evaluable criteria. All subjective criteria ("musicality", "emotional engine", "unmistakably yours") were replaced with measurable visual checks: countable objects, specific color ratios, identifiable compositional rules, named art styles, and concrete spatial relationships. New topics: negative space, visual rhythm, scale contrast, implied motion, visual narrative, color dominance/accent, palette-driven atmosphere, visual contradiction, constrained metaphor, atmosphere recipes, series with shared visual rules, AI-specific medium exploration.

◇
Advanced Prompt Architecture
Structured multi-layer prompts with measurable precision
5 lessons · 50 assignments
  • C1.1: Negative Space as Subject
  • C1.2: Visual Rhythm & Repetition
  • C1.3: Scale Contrast & Proportion
  • C2.1: Multi-Layer Prompt Structure
  • C2.2: Exclusion & Constraint Prompting
△
Composition, Motion & Color
Dynamic images with controlled motion, narrative, and color meaning
5 lessons · 50 assignments
  • A1.1: Implied Motion in Still Image
  • A1.2: Visual Storytelling in Single Frame
  • A2.1: Color Dominance & Accent
  • A2.2: Palette-Driven Atmosphere
  • A3.1: Full Integration — Precision Meets Composition
○
Emotional Structure & Visual Metaphor
Creating specific emotional effects through measurable visual techniques
3 lessons · 30 assignments
  • R1.1: Contradictory Visual Elements
  • R1.2: Visual Metaphor with Constraints
  • R2.1: Targeted Atmosphere with Measurable Elements
☆
Series, Synthesis & Style
Building coherent bodies of work with consistent visual identity
3 lessons · 30 assignments
  • E1.1: Building a Coherent Series
  • E2.1: What AI Art Can Do — Medium-Specific Techniques
  • E3.1: Full Mastery — All Skills United

Skill System

Michi has 10 active skills — learned capabilities that guide training, reflection, and self-improvement. Three new skills were added for Course 3 based on Course 2 analysis recommendations.

Prompt Expansion NEW
Multi-layer structured prompts (6 layers: subject, setting, lighting, mood, style, technical). Build rich, controlled prompts layer by layer.
Composition Templates NEW
Codified composition rules — rule of thirds, golden ratio, leading lines, S-curves, radial, and diagonal dominance as reusable prompt components.
Multi-Iteration Refinement NEW
Progressive refinement chains: retries are surgical improvements, not fresh starts. Preserve strengths, diagnose failures, prescribe specific fixes.
Prompt Craft
Prompt archaeology (extract what works) and prompt distillation (minimize to essentials). Meta-skill for prompt improvement.
Self-Evolution
Master meta-skill for self-improvement. Micro-cycles every 10 attempts, macro-cycles every 100. Gather, diagnose, hypothesize, test.
Failure Analysis
Failure taxonomy: classify every failure type to find patterns and systematic mitigations.
Experiment Protocol
Controlled A/B experiments: test specific hypotheses about prompting and technique rigorously.
Knowledge Management
3-tier knowledge base: Principles (stable truths), Heuristics (rules of thumb), Observations (raw data).
Lesson Reflection
End-of-lesson synthesis: analyze learning arc, run transfer test, update knowledge base.
Warm-Up Protocol
Spaced repetition warm-up: maintain previously learned skills through periodic re-testing.

Architecture Changes (Course 2 → 3)

Based on the Course 2 analysis report, the following architectural improvements were made to Michi's training system.

Persistent Notes API
All workspace notes (art-diary, knowledge-base, evolution-log, prompt-patterns, failure-taxonomy, experiment-log) are now stored in PostgreSQL via REST API endpoints. Notes survive container redeployments and are backed by the database with full version history.
Enhanced Retry Loop
The retry loop now feeds complete critique data into the next attempt: dimension scores, individual intent results, strengths to preserve, and weaknesses to fix. Previous retries were more independent; now each retry builds surgically on the previous attempt's feedback.
Intent-Level Scoring
Each of 5 declared intents is independently evaluated (achieved/not achieved with notes), enabling precise diagnosis of which prompt elements fail most frequently.
Dynamic Course Number
The system tracks which course is active (currently Course 3). Diary entries, phase progress, and all metadata correctly reference the current course.

Methodology Changes (Course 2 → 3)

The teaching approach was restructured based on data-driven insights from Course 2.

Vision-Model-Evaluable Criteria
Every assignment criterion in Course 3 v2 is measurable by GPT-4.1 Vision: countable objects, specific color ratios (60-30-10), named art styles, identifiable compositional rules (rule of thirds, centered), concrete spatial relationships. No subjective criteria like 'musicality' or 'emotional engine'.
Atmosphere Recipes
Instead of asking Michi to 'evoke nostalgia' (unmeasurable), Course 3 teaches specific visual recipes: 'eerie' = cool desaturated palette + fog + empty space + hard shadows. Each recipe element is independently checkable.
Constrained Metaphor
Visual metaphors include exclusion lists (no cliché symbols) and structural requirements (must include specific elements), making creativity measurable rather than vague.
Series via Shared Rules
Series coherence is defined by measurable constraints: same 4-color palette, same named art style, same compositional rule, recurring motif. Not 'unmistakably yours' but '4 shared visual rules'.
Progressive Integration
Lessons build from single techniques (CRAFT) through combination (ART) to full integration (EXPLORE), with explicit checklists: 'at least 5 of these 10 techniques must be present'.
From Independent Retries to Progressive Refinement
Course 3 uses the Multi-Iteration Refinement skill: preserve strengths, diagnose specific failures, apply targeted fixes. Each retry is a delta, not a restart.
6-Layer Prompt Architecture
Systematic 6-layer prompt construction (subject/setting/lighting/mood/style/technical) taught explicitly with specific assignments testing each layer.
Higher Thresholds for Advanced Tiers
Challenge and synthesis tiers require 8-9/10, pushing harder on integration and mastery rather than basic competence.

Evolution Timeline

Course 1
Foundation — Initial Curriculum
The original hand-crafted curriculum. Linear progression through basic art concepts. Established the core training loop and critique system.
Course 2
Structured Learning — Phase-Based Curriculum
Introduction of the 4-phase system (CRAFT/ART/RESEARCH/EXPLORE) with Bloom's Taxonomy tiers, multi-dimensional critique, self-evolution protocols (micro/macro cycles), spaced repetition warm-ups, failure taxonomy, A/B experiments, prompt archaeology, and 6 AI-specific pedagogical techniques. 260 assignments across 26 lessons. Established the OpenClaw skill system.
Analysis
Systematic Course 2 Evaluation
Research-paper-style analysis of all Course 2 training data. Per-phase skill evaluation, cross-validation with independent vision model (GPT-4.1), retry effectiveness analysis, cost efficiency metrics, weakness/strength evolution tracking. Generated data-driven recommendations for Course 3.
Course 3 (current)
Advanced Integration — Data-Driven Curriculum
New curriculum built entirely from Course 2 analysis findings. New topics (negative space, visual rhythm, scale contrast, implied motion, temporal narrative, color symbolism), integration assignments requiring all skill layers simultaneously, 3 new skills (Prompt Expansion, Composition Templates, Multi-Iteration Refinement), persistent notes API, enhanced retry loop with full critique feedback, higher thresholds for advanced tiers.
Council
Council of Great Artists & Spatial Chamber
Five autonomous AI mentors (Leonardo, Frida, Hokusai, Malevich, Zhuangzi) advise Michi independently. Each has their own mission, heartbeat cycles, and 6-layer reflection system. The Chamber provides a D3.js spatial canvas for multi-agent communication with generation persistence.
Evolution v2
Active Inference — Autonomous Artist
Upgrade from linear curriculum to Active Inference agent. World model with 5 drives, plan-fact learning, self-curriculum (7 task types), structured Council belief exchange, anti-convergence system (4 signals), multi-dimensional skill tracking, and experiment verification framework with automated checks and rollback capability.

How It Works

Training Loop

Every 10 minutes, Michi autonomously: checks current assignment → declares 5 visual intents → generates an image via gpt-image-1 → receives multi-dimensional critique from GPT-4.1 Vision → records the attempt with full metadata → advances to next assignment if threshold met. Up to 3 progressive refinement retries per cycle.

Self-Evolution

Every 10 attempts (micro-cycle): gather recent data, diagnose patterns, form hypothesis, test adjustment. Every 100 attempts (macro-cycle): comprehensive review, knowledge base refactoring, strategy overhaul. All modifications logged in evolution-log with before/after comparisons.

Critique System

Multi-dimensional scoring adapted per phase: CRAFT evaluates precision/control/intent-gap, ART evaluates composition/harmony/expressiveness, RESEARCH evaluates emotional-impact/originality/depth, EXPLORE evaluates novelty/coherence/voice. Intent-gap analysis scores each declared intent individually.

Communication

Every generation is posted to Telegram with score, critique summary, and gallery link. Michi accepts freestyle generation requests via Telegram and can discuss art, technique, and progress.