Four authors

Four ways to ship

The same era, the same AI-assisted tooling, four different philosophies — one optimizing for velocity, one for institutional continuity, one for measured balance, and one that inherited a format and re-skinned it.

craig7351 (Book AI)

Solo developer + AI pair

Five 3D browser games shipped in 36 days (2026-06-12 → 07-18), every one deployed live with a real Cloudflare D1 backend — leaderboards, message boards, online-presence charts. 248 commits, ~29k lines of hand-written code, and effectively 100% of commits co-authored by Claude (Opus 4.8 for the first four, Fable 5 for the fifth).

The signature workflow: a human directing design, verifying on real devices and handling real-world constraints (free-tier quotas, iOS quirks), while the AI produces code at burst cadence — 63 commits in a single day at peak. The repos double as artifacts of this workflow: reskin playbooks, READMEs written for 'the next AI', a custom Claude skill, even a prompt-injection incident report.

Consistent strengths: completeness and game feel, plus unusually mature free-tier cost engineering. Consistent debts: zero-to-minimal testing, god-file architecture, and a recurring weak-password signature ('0501') that leaks into public repos.

6.3avg total
5projects
248commits
28.9klines

Open Culture Foundation (OCF)

Nonprofit organization

One project maintained over 22 months (2024-09 → 2026-07): a digital-security curriculum for Taiwanese civil-society organizations, run as institutional infrastructure — shared org account, PR reviews with issue references, issue templates, deploy-branch strategy, scoped tokens, CC-BY 4.0 licensing.

Where the solo author optimizes for velocity, OCF optimizes for continuity and audience fit: written editorial rules, a policy template mapped to Taiwanese NGO governance, Tor access for at-risk users — and a late-2026 burst that quietly added a 7,600-line 3D teaching game.

7.2avg total
1projects
160commits
7.6klines

Leo Chang · 最先生 (LeoGGcat)

Solo creator + collaborator + AI

One game, 18 consecutive days, 267 commits and 32 released versions (2026-08-07 → 08-24) — a text-only Taiwanese baseball career sim shipped on its own domain as an installable PWA. The workflow is a hybrid no other author here uses: the owner (credited in-game as 最先生 Mr.TheMost) directs design, a second developer contributes through pull requests, and Claude is the commit author on 39 commits and co-signs 30 more.

The signature is measurement. Balance is not tuned by feel but re-derived from full-flow Monte Carlo runs — thousands of simulated careers used first to prove the old Hall-of-Fame thresholds were mathematically unreachable, then to re-anchor the whole ladder to explicit population targets. The code carries the same habit: comments justify decisions with numbers (contrast ratios, measured season counts, surgery frequencies) instead of restating what the line does.

Consistent strengths: domain-modelling depth, comment quality and product-grade finish (four themes, PWA, share card, wiki, changelog). Consistent debts: no tests, no CI, a calibration harness that never entered the repository, and 173 hand-maintained cache-bust tokens standing in for a build step.

7.0avg total
1projects
267commits
8.7klines

jarvanthevoyager

Solo producer, borrowed format

One game in a single 3,500-line HTML file, published through the GitHub web UI in 17 commits across four days — 'Add files via upload', 'Update index.html'. That history is therefore not a development history at all: the work happened off-platform, and the artifact is the only evidence of how it was built.

The README credits the original concept to 最先生 (@mr.themost) — the designer behind YaKyoLife. This is a format handed to a second producer: the same seeded-life chassis re-skinned onto cheer-squad careers, a subject with no obvious prior example. What the producer adds is output discipline (consistent escaping, cookieless analytics, a documented storage table) and the audit's only shipped integrity checker.

The trade-off is legibility and provenance: one global scope holding 131 top-level functions, documentation that has drifted from the build (stale counts, three broken links, a licence badged but absent), and balance argued by observation where the sibling project argues by simulation.

6.0avg total
1projects
17commits
3.5klines
Head to head

Author averages, overlaid

craig7351 = mean of five games; every other author = their single project. The overlaps tell the story: craig leads on completeness and game feel, OCF on content and security posture, LeoGGcat on design and code quality — and every one of them collapses on testing.

Technical DepthGame DesignBackend & SecurityCode QualityTestingCompletenessOriginality
craig7351 (Book AI)Open Culture Foundation (OCF)Leo Chang · 最先生 (LeoGGcat)jarvanthevoyager
craig7351 · timeline

Five games in 36 days

The arc of a solo-plus-AI sprint — and where the skill curve shows.