Project Timeline

51 days of
building.

A chronological record of how a 10-issue educational comic series was built entirely by human-AI collaboration using multi-agent orchestration.

51
Days
36
Sessions
702
Agents Spawned
2,133
Images Generated
All Sessions
01 — Session Log

36 sessions across
51 days.

From the first 6-hour marathon that produced all 10 issue drafts to the 7-day mega-session that reviewed every generated image.

# Date Duration Agents Focus
1Feb 7~6h37Research + write all 10 issues
2Feb 7~2h4Layout design continuation
3Feb 7~4h31Proofreading + editorial review (Issues 1-2)
4Feb 8~6h21Proofreading + fixes (Issues 3-5)
5Feb 8~12h51Full pipeline: edit, red team, proofread, fix (Issues 5-10)
6Feb 9~4h19Final editorial fixes (Issue 9 etc.)
7Feb 9brief1Spot fix
8Feb 12~8h42SVG visual review (18 visual reviewers) + interactive labs
9Feb 14brief4Small fixes
10Feb 15~6h53V2 rewrite: 10 writers + 10 editors + 10 red team
11Feb 15~12h83V2 continuation: more writers + editors + red team
12Feb 15~3h16V2 assembly and fixes
13Feb 15~1h9More fixes
14Feb 15brief4Spot fixes
15Feb 15brief8Spot fixes
16Feb 15~2h14Layout review
17Feb 15brief3Small review
18Feb 15brief2Spot fix
19Feb 15~4h16Layout/assembly work
20Feb 15brief1Single task
21Feb 15brief1Single task
22Feb 16~6h39Visual review round (20 visual reviewers)
23Feb 16~4h19More visual review
24Feb 16~6h26Visual review + frontend polish
25Feb 16brief1Single task
26Feb 16-237 days62Image review + regeneration (438 images across 10 issues)
27Feb 282.5h9Git commits + process analysis of 50 sessions
28Feb 28<1m1Micro session: repo improvement suggestion
29Feb 2818m7Repo improvements + execution
30Feb 28 - Mar 76+ days1Code quality refactoring (12 phases), CI, tests
31Mar 17-214 days15Character design + style experiments + FLUX.2-dev on DGX Spark
32Mar 21-2219h14Photo-real Scribble, multi-image API, 500+ character designs
33Mar 258h34Experiment reorg, documentary, behind-the-scenes
34Mar 26~10h40Repo restructure, two-repo split, Future Canvas (160 concept art), pipeline hardening, narrative guidelines
35Mar 28~8h6Concept production pipeline v3/v4 — creative memory, story sequences, VLM critique, PickScore
36Mar 29~10h8Deep research (61 breakthroughs, 98 ref images), science educator (27 threads), concept production v5
Chronological Phases
02 — The Build

Ten phases from
genesis to vision.

The project moved through distinct phases: research, editorial review, image generation, quality review, character design, documentation, reorganization, and autonomous production.

Phase 1 — Feb 7, 2026
Genesis
Sessions 1-3 · 72 subagents
Built the entire 10-issue series from scratch in a single day.
  • 10 Researcher agents spawned in parallel for historical facts
  • 10 Writer agents drafted page-by-page scripts with dialogue and narration
  • 5 Layout Designer agents created panel manifest JSON files
  • Proofreading and editorial review began immediately on completed issues
  • The "Constellation Pattern" established: agents communicate only through files
All 10 issue scripts drafted in ~6 hours
Phase 2 — Feb 7-9, 2026
Editorial Gauntlet
Sessions 3-7 · 122 subagents
Multi-round editorial review and quality fixes across all issues.
  • Editor agents checked factual accuracy, pacing, and clarity
  • Red Team agents challenged for errors, oversimplifications, and bias
  • Proofreader agents verified SVG rendering, robot consistency, link validation
  • Fixer agents applied patches based on review reports
  • Multiple proofreading rounds per issue (up to v4 for some)
All 10 scripts approved through editorial pipeline
Phase 3 — Feb 12, 2026
Interactive Labs & Illustrations
Session 8 · 42 subagents
Built supplementary interactive content and SVG visual quality review.
  • 18 Visual Reviewer agents examined SVG diagrams across all issues
  • Built 16 interactive JavaScript labs (Turing Machine sandbox, neural network playground, sorting arena)
  • Created 10 detailed illustrations
Supplementary educational content complete
Phase 4 — Feb 15, 2026
V2 Scripts
Sessions 10-21 · ~210 subagents across 12 sessions
Complete rewrite of all 10 issue scripts for V2. The most agent-intensive day of the project.
  • Massive parallel spawning: 10 writers + 10 editors + 10 red team simultaneously
  • Multiple rounds of revisions with fixer agents
  • Layout review across batches of issues
  • Up to 83 subagents in a single session
  • 12 separate sessions as the user pushed to complete V2 scripts in one day
V2 scripts complete
Phase 5 — Feb 16, 2026
Image Generation & Visual Review
Sessions 22-25 · 85 subagents
AI image generation using Z-Image-Turbo (local MPS) and visual quality review.
  • Generated 438 images across 10 issues using Z-Image-Turbo
  • 20+ Visual Reviewer agents checked rendering quality in batches
  • Frontend polish: accessibility, reduced-motion, design standardization
  • SVG sanitization pass across all issues
V1 published — inline SVG robots, no AI images
Phase 6 — Feb 16-23, 2026
The Mega-Session
Session 26 · 62 subagents · 63.5MB transcript
Systematic review of all 437 AI-generated images and targeted regeneration over 7 days.
  • 10 parallel Image Reviewer agents (one per issue), each examining 32-58 images
  • Findings: 438 images reviewed — 330 OK, 3 critical, 71 warnings, 34 informational
  • Dominant issue: AI text garbling in speech bubbles
  • Secondary: anachronistic hardware (modern computers in 1960s scenes)
  • Regenerated 107 images with fixed prompts across all 10 issues
  • Post-regen review: 10 more agents verified fixes
V2 published — all images reviewed and fixed
Phase 7 — Feb 28, 2026
Process Retrospective
Sessions 27-30 · 18 subagents
Git housekeeping, process analysis, and code quality refactoring.
  • Organized 195 uncommitted files into 5 logical git commits
  • 4 analysis agents mined 420MB of session transcripts for process improvements
  • Major code quality refactoring: 12 phases, 44 tests, GitHub Actions CI
  • Caught a real bug: batch_generate.py referenced a removed variable
  • 22 production rules codified from 50-session analysis
CLAUDE.md hardened with 22 production rules
Phase 8 — Mar 17-22, 2026
Character Design & Style Exploration
Sessions 31-32 · 29 subagents
Exploring visual styles and character designs using FLUX.2-dev on DGX Spark GPU.
  • Connected to Sparky API (DGX Spark running FLUX.2-dev NVFP4)
  • Generated 769 character design images (concepts, era-themed, material studies)
  • Produced 8 full art style experiments for Issue 1 (484 images)
  • Extended Sparky API to support multi-image editing
  • Leading character emerged: Scribble v2 "The Cruiser"
8 visual style experiments + 769 character designs
Phase 9 — Mar 25, 2026
Organization & Documentary
Session 33 · 34 subagents
Renamed experiments, built documentary site, organized the asset repo.
  • Renamed v3/v4 directories to descriptive experiment names
  • Built behind-the-scenes documentary with pipeline visualization
  • Updated all script references and internal paths
  • Session timeline extraction and documentation
Project transitioned from production to reflection
Phase 10 — Mar 26, 2026
Reorganization & Vision
Session 34 · ~40 subagents
Major repo restructure, philosophical foundation, and 160-image concept art generation for the Future Canvas.
  • Split into two repos: workflow (scripts/agents) and assets (hexley-dev.github.io)
  • Issue versioning introduced: v1/, v2/, staging/ directory structure
  • Archived 20 experiment scripts to archive/experiments/ and archive/migrations/
  • Git identity rewrite: all commits now attributed to Hexley
  • Documentary site expanded: 7 pages with shared nav/footer, two-tier navigation
  • "Why We Build" philosophical foundation page and 10 narrative guidelines codified
  • "On Our Mind" living ideas page added to Latest nav
  • Future Canvas: 160 concept art prompts (60 robots, 60 scenes, 40 styles) generated via Sparky API in 169 minutes
  • Pipeline hardening: validate_schema.py, image_gen_audit.py, pipeline_status.py, generate_concept_art.py
Two-repo architecture + 160 concept art pieces + philosophical foundation
Phase 11 — Mar 28-29, 2026
Autonomous Concept Production
Sessions 35-36 · ~14 subagents
Built a self-improving concept art pipeline grounded in real scientific breakthroughs, evolving from standalone images to research-backed narrative sequences.
  • concept_production_v3: creative memory system with explore/exploit/pivot modes, structural style enforcement — 51 kept images from 74 generated
  • concept_production_v4: 3-beat story sequences (before/moment/after), VLM visual critique via Claude CLI, always-refine via image-to-image editing, CREA Critic pattern, PickScore + HPSv2 quality scoring
  • deep_research_sweep: fetched 40+ real sources (DeepMind, NASA, Nature, Anthropic) extracting 61 breakthroughs with visual language fields
  • deep_research_enhance: followed links from hub pages, fetched and analyzed 98 scientific reference images with Claude vision
  • science_educator: connected breakthroughs into 27 narrative threads with 3-beat visual arcs — beat 2 = "the science itself as visual spectacle"
  • concept_production_v5: draws from breakthrough database + connection threads, 25K+ chars of visual context per round, thread mode every 3rd round
  • New CLAUDE.md rules codified: Never Kick the Can (#24), Surface Depth Questions (#26), SOTA verification (#17)
6 new scripts + 61 breakthroughs + 282 concept art images (v3-v5)
By the Numbers
03 — Statistics

The production
in numbers.

691
MB Transcripts
63.5
MB Largest Session
83
Peak Agents / Session
107
Images Regenerated

Images Generated

CategoryCount
V2 published (10 issues)438
Style experiments (8 styles)484
Character designs769
Future Canvas concept art (v1)160
Concept production (v3/v4/v5)282
Total2,133

Agent Roles

RolePurposeEst. Spawns
Visual ReviewerScreenshot-based rendering review~80
ProofreaderRendering verification~60
WriterNarrative scripts~40
EditorAccuracy + pacing review~40
Red TeamAdversarial review~30
FixerMechanical patch application~30
ResearcherHistorical fact gathering~20
Layout DesignerPanel manifest creation~20
Image ReviewerAI image quality review~20
AnalysisProcess mining~8

Git Milestones

DateCommitSummary
Feb 9413f97dInitial commit: 10-issue series
Feb 1291eabd316 interactive labs + 10 illustrations
Feb 1662c4174Frontend polish + accessibility
Feb 16b188821Complete pipeline: all images + SVGs + HTML
Feb 28b0579dbImage review tooling
Feb 28d10d203Prompt fixes from review
Feb 2885a311cRegenerate 107 images
Feb 28078aca5Reassemble all 10 issues
Feb 28c266393Review reports + regen specs
Feb 283b47323Process review (50-session analysis)
Feb 280375179User prompting analysis
Feb 284b61b35Code quality + tests + CI
Feb 28bfc28b6CI fix
Mar 21803c296Behind-the-scenes documentary
The Story
04 — Narrative Arc

How the project
unfolded.

From a blank repository to a published comic series with 2,133 AI-generated images, told in eleven turning points.

1
Day 1 — Feb 7 The entire 10-issue series was researched and written in a single session using 37 parallel agents. The Constellation Pattern was established from the start: agents communicating only through files, never through the orchestrator's context.
2
Days 1-3 — Feb 7-9 Intensive editorial gauntlet. Editor and Red Team agents challenged every issue. Proofreaders verified rendering. Multiple revision rounds per issue. The pipeline crystallized: content must be locked before committing to image generation.
3
Day 6 — Feb 12 The project expanded beyond comics into a comprehensive educational platform with 16 interactive JavaScript labs and 10 detailed illustrations.
4
Day 9 — Feb 15 The most intense day. 12 sessions, 210+ subagents. Complete V2 rewrite of all scripts with simultaneous writer + editor + red team for all 10 issues.
5
Day 10 — Feb 16 Image generation day. 438 images produced with Z-Image-Turbo. V1 published. But the real work was just beginning: the image review pipeline had not been built yet.
6
Days 10-17 — Feb 16-23 The mega-session. A 63.5MB transcript spanning 7 days. 438 images reviewed by 10 parallel vision agents, 107 regenerated with fixed prompts. The most expensive lesson: AI image models garble text in speech bubbles.
7
Day 22 — Feb 28 Retrospective day. Process analysis of 50 sessions yielded 22 production rules. Code quality refactoring added tests, CI, and caught a real bug.
8
Days 39-45 — Mar 17-22 The style exploration era. FLUX.2-dev on a DGX Spark GPU opened new possibilities. 8 visual style experiments, 769 character designs, photo-real brass robots. Scribble v2 "The Cruiser" emerged as the leading character.
9
Day 47 — Mar 25 Housekeeping and documentation. Experiments renamed, documentary built, timeline extracted. The project transitioned from production to reflection.
10
Day 48 — Mar 26 The great reorganization. The project split into two repos (workflow vs. assets), issue versioning was introduced, and 20 experiment scripts were archived. The documentary site grew to 7 pages with a philosophical foundation ("Why We Build") and 10 narrative guidelines. Three creative director agents designed 160 concept art prompts for the Future Canvas — robots, future scenes, and art styles — all generated via Sparky API in under 3 hours. Four new pipeline tools hardened the production workflow. The project found its voice.
11
Days 50-51 — Mar 28-29 The pipeline learned to see and to read. Concept production evolved across three generations: v3 added creative memory and style enforcement, v4 introduced 3-beat story sequences with VLM visual critique (Claude evaluating its own generated images) and automated quality scoring. Then the pipeline went deeper — a research sweep harvested 61 real breakthroughs from 40+ scientific sources, a vision agent analyzed 98 reference images, and a science educator wove them into 27 narrative threads. V5 drew from all of it: 25,000 characters of visual language context per generation round, grounded in real science rather than imagination alone. Six new scripts, three new CLAUDE.md rules, and the concept art count climbed to 442.
Takeaways
05 — Lessons Codified

10 rules from
the trenches.

Rule 01
One GPU process at a time
A 24GB GPU can only run one diffusion model. Launching two crashes both.
Rule 02
Never bulk-delete before confirming replacements work
The session that deleted 61 images before verifying torch was functional was the costliest mistake.
Rule 03
Verify long-running tasks early
"Running" does not mean "working." Check output after the first iteration.
Rule 04
Fix patterns, not instances
One agent fixing all occurrences of an error class prevents 20-spawn proofreading loops.
Rule 05
Content review before image generation
Scripts are cheap to iterate; images are expensive. Lock content first.
Rule 06
Historical figures need reference photos
Without them, AI generates generic faces that do not match real people.
Rule 07
AI image models garble text
Speech bubble text should be kept under 12 words and use simple vocabulary.
Rule 08
Monospace fonts for all SVG text
Cursive fonts cause cramming and overlap in rendered SVG diagrams.
Rule 09
Save state obsessively
Update MEMORY.md every 3 interactions. The orchestrator session is a single point of failure.
Rule 10
Rules must live in files, not conversation
If a mistake can happen once, it can happen again. Codify immediately.