JingWen Fan

May 28, 2026 · View on GitHub

Generated from local Claude Code + Codex + Cursor + Antigravity data across 148 active days · 2026-05-28 Personal philosophy: decompose complex problems into irreducible facts or constraints, then reason only from those undeniable ground truths rather than experience, analogy, or inherited conclusions.


At a glance

  • Across 148 active days: 605 Claude sessions + 2,094 Codex threads + 105 Antigravity tasks, with 177,932 Claude/Codex messages
  • Daily output: 29 local commits / 5,776 lines of churn / 10 GitHub contributions
  • Active across 13 local git repos, spanning 7 languages (Python / TypeScript / Rust / Vue / HTML / JavaScript / Jupyter)
  • Same-period GitHub: 3,467 commits / 114 PRs / 38 issues / 3,706 total contributions (85 repos owned · 294 PRs lifetime · 46 issues lifetime)
  • AI input: 429.9M Claude paid/new tokens + 11.56B Codex tokens; 10.22B Claude cache-read tokens reused (Cache leverage 23.77× — every new token pulls ~24 cached tokens back in)
  • Main tools: Claude Code (Opus 4.6 + Sonnet 4.6 + Opus 4.7-1M + Haiku 4.5) + Codex CLI (GPT-5.5 + GPT-5.4) + Cursor (composer) + Antigravity (March cohort)

🚀 Velocity & Leverage — AI gives one person small-team delivery capacity

13 local repos, 7 languages, 29 commits/day, 4,292 commits and +735K lines added in half a year — this kind of cross-stack breadth and delivery density only works when Claude + Codex are running as the primary IDE.

MetricValueNotes
Daily commits294,292 local commits / 148 active days
Daily code churn5,776 lines(+735K -120K) / 148
Active repos13Candidate repos with commits by the current git author in the past year
Cross-stack languages7Python / TypeScript / Rust / Vue / HTML / JavaScript / Jupyter
GitHub burst window2026-03-13 → 2026-03-19Peak day 12,301 Claude messages (3-18)
Open-source reach~2,217 starsAcross 14 starred public repos; flagship 1,246★ (antigravity-workspace-template)

🤖 AI-Native practice

This is not occasional AI usage. Four AI coding tools running in parallel, divided by task type.

Multi-model / multi-tool orchestration

Model / Toolsessions / threadsspent tokenscache-readleverageTypical role
Claude Opus 4.6(main driver)308.9M7.79B25.27×Deep reasoning, long-chain planning
Claude Opus 4.7 (1M)(long-context driver)83.3M1.82B21.85×1M-context review, cross-file audits
Claude Sonnet 4.6(lightweight)12.3M357.8M29.12× (most fluent)Rapid iteration, batch tasks
Claude Haiku 4.5(fast-lane)25.4M250.6M9.88×Short queries, command classification
GPT-5.5 (Codex)1,108 threads6.31BPrimary reasoning lane, cross-tool diagnosis
GPT-5.4 (Codex)810 threads4.30BSecondary lane
GPT-5.4-mini / 5.3-codex / 5.2 / 5.1-codex-max39 threads379MExperiments / fallback
Cursor (Anysphere)7 workspacescloud-authoritativeComposer / inline IDE editing
Antigravity (Google)105 tasksTask-based planning / walkthrough

Total orchestration breadth = 4 AI tools live in parallel + 9 distinct model versions. Inside the Anthropic stack: Opus (deep) + Sonnet (fast) + Haiku (snappy) + 1M-context, all four. Inside the OpenAI stack: GPT-5.5 → 5.4 → 5.3 → 5.2 → 5.1 across multiple generations. This is the extreme form of "multi-engine orchestrator" — not just having access, but routing tasks to the right lane reflexively.

Power-feature usage

  • Plan-mode: /plan used 96 times (used at decision points; 1.7% entry rate on plain prompts — selective, not reflexive)
  • Reasoning effort control: /effort used 1,198 times — absolute dominant command, 59.8% of all slash usage; switching reasoning depth has become a hardened muscle memory
  • Skills: 17 total (Claude 6 installed + Codex 11; includes 4 self-built: chronicle / codex-primary-runtime / playwright / sora)
  • Plans: 11 (centered on billing / payment / router channel management / trace_id / captcha integration)
  • Hooks: 1 (Claude) · Codex Automations: 4 (daily AI sessions, fully scheduled)
  • /btw: 62 times (custom slash command) · /vibe-forge: 20 times (self-built skill invocation)

Prompt caching maturity

Claude paid/new tokens total 429.9M, while cache-read totals 10.22B. Every new Claude token pulls roughly 24 cached tokens back into play; cache-read is 95.96% of Claude I/O. Sonnet 4.6 is the most fluent cacher per dollar (29.12× leverage); Opus 4.6 carries 71.86% of total spend as the workhorse.

Reasoning effort preference (Codex)

effortcountshare
xhigh1,57275.1%
medium27112.9%
unspecified1386.6%
high874.2%
low261.2%

75% of Codex threads run on xhigh — not "occasionally deep", but "almost always ask AI to go full strength". This effort profile sits in the top tail of the ecosystem's power-user distribution.

🔧 AI infrastructure — not just using AI, but building tooling for AI

From skills to hooks to automations, from plugins to launchers, a personal infrastructure has been built so AI can work harder on the user's behalf.

Self-built skills / automations

NameDescriptionTool
vibe-forgeCustom workflow forgeClaude (slash 20× · GitHub: 14★ Rust)
/btwOut-of-context instruction shortcutClaude (62 uses)
chronicleTimeline / log aggregationCodex
codex-primary-runtimeCodex primary runtime skillCodex
playwrightBrowser automationCodex
sora(Sora video?)Codex
avcAgent View Controller (own open-source · 46★ JS)Claude + Codex
chronicle / daily-ai / ai-21-00-ai-session-004 Codex automationsCodex (scheduled / manual)

Installed skills (within the 17 total)

  • Claude (6): avc · docx · pdf · pdf-reading · pptx · xlsx
  • Codex (11): avc · chronicle · codex-primary-runtime · doc · docx · pdf · pdf-reading · playwright · pptx · sora · xlsx
  • Cross-tool reusable: avc / docx / pdf / pdf-reading / pptx / xlsx — 6 skills live in both Claude and Codex simultaneously, demonstrating real AI-infrastructure interoperability

Other infrastructure

  • Hooks: 1 (Claude)
  • Codex automations: 4 (chronicle / daily-ai / 21:00 auto-session / manual trigger)
  • Cross-platform plugin: Readme.skill itself (this report's generator), 124★, supports 6 tools

🛠️ AI collaboration style

Top 10 slash commands

#CommandCountMeaning
1/effort1,198Switch reasoning depth (absolute leader)
2/usage137Check tokens / quota
3/plan96Enter plan-mode
4/resume87Resume session
5/clear66Clear context
6/btw (custom)62Out-of-context note
7/rate-limit-options29Rate-limit handling
8/compact29Compact context
9/plugin22Plugin management
10/vibe-forge (custom)20Invoke own forge skill

Session architecture

  • Typical flow: Enter session → /effort (almost every session) → /plan decision-point → deep iteration → /compact or /clear reclamation → /resume to continue across sessions
  • 5,642 plain prompts vs 2,005 slash commands (plain 73.8% / cmd 26.2%) — command-assist rate above 1/4, far above beginner range
  • 96 /plan entries (1.7% of plain prompts) — used at decision points, not reflexively
  • /effort 1,198 times + 66 /clear + 29 /compact — active context management
  • /resume 87 times → strong session continuity across devices / time windows
  • Average depth: 294 messages / session (177,932 / 605) — most are long-form sessions
  • Longest session: 172.8 hours / 1,444 messages (one 7-day continuous context)

📂 Projects & domain distribution

Active across 77 Claude projects + 15+ Codex projects, deduplicated to Top 12.

DomainProject countSignature
Product backend / full-stack6LLM router channel management / billing / payment / mailer / CMO backend / user service
Product frontend3dashboard / admin console / router frontend (React + TypeScript)
AI tools / skill5Readme.skill / vibe-forge / AVC / antigravity-workspace-template / easy-claude-code
Data / analytics1Industry data crawl / user profiling (Python + Codex scheduling)
ML / RL / research1post-training / grpo-rlvr-llm-training

Top projects (anonymized)

ProjectClaude sessionsCodex threadsCursorAntigravityGit commitsOrchestrationDomain
Project A1,1862,851Claude-dominatedBackend / data crawl (Python)
Project B141406487Multi-engine [Cu]LLM router channel management (TS + Python)
Project C457NO_GITCodex-dominatedExternal codebase audit
Project D2976(strong March correlation, 105 tasks)157Multi-engine [Cu]AI tool template (open-source 1,246★ Python)
Project E14148344Codex-dominatedCMO backend service (Python)
Project F368011Dual-engineCLI tool
Project G109377Multi-engine [Cu]Dashboard / admin
Project H8(low)130git-dominatedBusiness frontend
Project I25(low)36Claude-dominatedTypeScript full-stack
Project J1735Claude-dominatedMailer agent (Python)
Project K9(low)100git-dominatedRouter frontend
Project L133NO_GITCodex-dominatedIndustry data crawl (Codex overnight scheduled)

Orchestration distribution: 3 multi-engine · 1 dual-engine · 4 Claude-dominated · 3 Codex-dominated · 2 git/single-tool

Multi-engine battlegrounds: Projects B / D / G run Claude + Codex + Cursor in parallel — daytime Claude for long reasoning, late-night Codex on xhigh for implementation, Cursor for in-workspace editing. Three shifts, one developer.

🧬 Evolution curve — AI usage is evolving

2026-01  Codex starts: 29 threads / 38.6M tokens / CLI 0.118.0
2026-02  Daily-ization: 23 threads / 251M tokens (warm-up)
2026-03  ⚡ Claude Code joins, mass explosion: 261 sessions / 114K msgs / peak day 3-18 (12,301)
         Antigravity concentrated experiment (3-17 → 3-31, 105 tasks across 4 artifact types)
         Codex monthly leap: 132 threads / 702M tokens
2026-04  Two-engine maturity: Claude 282 sessions / Codex 948 threads / 6.16B tokens
         /effort muscle memory fully formed (60% of slash usage)
2026-05  Steady scaling: Codex 962 threads / 4.41B tokens; Claude usage shrinks to critical nodes
         Cursor enters as daily IDE / antigravity flagship project under continuous maintenance

Monthly activity trend

MonthClaude sessions / msgsCodex threads / tokensMilestone
2026-0129 / 38.6MCodex 0.118 launch
2026-0223 / 251MDaily-ization warm-up
2026-03261 / 114K132 / 702MMass explosion: Claude joins, Antigravity bursts, peak day 3-18
2026-04282 / 52.7K948 / 6.16BTwo-engine maturity, Codex takes off
2026-0562 / 11.2K962 / 4.41BCodex steady scale, Claude focuses on critical nodes

Model migration notes:

  • 2026-03 → 04: Codex spent ↑ 8.8× (702M → 6.16B), driven by gpt-5.5 / gpt-5.4 replacing the early gpt-5.1-codex-max / gpt-5.2-codex-spark cohort
  • 2026-04 → 05: Codex spent ↓ 28% (6.16B → 4.41B) but thread count +1.5% — per-thread efficiency improved (better caching, stable xhigh use)
  • Claude side: Opus 4.6 is the consistent main driver. From April, Opus 4.7 (1M context) joins to handle ultra-long-context scenarios (spent 83M), not replacing 4.6 but specializing. Sonnet 4.6 used for rapid iteration.

💡 Topics & keywords

router · billing · payment · agent · plan · skill · mcp · antigravity · vibe-forge · trace_id · plugin · dashboard · mailer · CLI · observability · captcha · channel management · billing multiplier · walkthrough · first-principles · LLM router · multi-tool orchestration · token economics · cache leverage · session architecture · sub-agent · plan-mode

Topics cluster in three rings:

  1. Commercial product: LLM router channel management / billing / payment / captcha / dashboard / user service
  2. AI tool construction: skill / vibe-forge / AVC / Readme.skill / antigravity template / mcp / plan / hook
  3. Observability & trust: trace_id full-link / log / observability / retry backoff / payment_sources

⏱️ Working rhythm

24h heatmap (Claude + Codex merged)

00  ▓▓                      Claude 38 · Codex 9
01  ▓                       Claude 16 · Codex 19
02  ▒                       Claude 6  · Codex 12
03  ▓▓▓▓▓▓                  Claude 1  · Codex 154  ← Codex overnight burst start
04  ▓▓▓▓                    Codex 99
05  ▓▓▓▓▓                   Codex 120
06  ▓▓▓▓▓                   Codex 130
07  ▓▓▓▓▓                   Codex 128
08  ▓▓▓▓▓▓▓▓                Claude 1  · Codex 189  ← morning peak
09  ▓▓▓▓▓▓                  Claude 5  · Codex 156
10  ▓▓▓▓▓▓▓▓                Claude 12 · Codex 182
11  ▓▓▓▓▓▓▓▓                Claude 60 · Codex 148  ← Claude morning peak
12  ▓▓▓▓▓▓                  Claude 24 · Codex 141
13  ▓▓▓▓▓▓▓                 Claude 47 · Codex 119
14  ▓▓▓▓                    Claude 45 · Codex 57
15  ▓▓▓▓▓▓                  Claude 49 · Codex 107
16  ▓▓▓▓▓▓▓▓                Claude 52 · Codex 148  ← afternoon double peak
17  ▓▓▓▓▓▓▓                 Claude 39 · Codex 153
18  ▓                       Claude 16 · Codex 19
19  ▓▓                      Claude 46
20  ▓▓                      Claude 55
21  ▓▓                      Claude 42
22  ▓                       Claude 23 · Codex 4
23  ▓                       Claude 28

Peak windows: Early 03-11 is a double peak from Codex automation + active morning work, afternoon 16-17 is the Codex second round, evening 19-21 is Claude wrap-up. Codex's heavy 3-7 AM activity correlates with the ai-21-00-ai-session-00 automation — AI doing the night shift.

Time span

  • First active: Codex 2026-01-01 · Claude Code 2026-03-09
  • Most recent: 2026-05-28 (today) · still active
  • Span: 148 days
  • Active days: 57 (Claude dailyActivity basis)
  • Peak day: 12,301 Claude messages (2026-03-18); top 5 highest-volume days all within the 3-13 → 3-19 week

💎 Token economics

440M new paid tokens leveraged 10.22B cache replay, ratio 1 : 24; total through Claude alone 10.65B, plus Codex 11.56B — full-stack ~22.21B tokens through. Equivalent to processing 15.5 billion Chinese characters with AI across 148 days — roughly 21,000 times the length of Dream of the Red Chamber, or 38,000 times War & Peace.

Per-model token breakdown (sorted by spent)

Modelspentcache-readleverage% of total spent
Claude Opus 4.6308.9M7.79B25.27×71.86%
GPT-5.5 (Codex)6.31B
GPT-5.4 (Codex)4.30B
Codex (unspec)606M
Claude Opus 4.7 (1M)83.3M1.82B21.85×19.38%
GPT-5.4-mini (Codex)265M
Claude Haiku 4.525.4M250.6M9.88×5.91%
Claude Sonnet 4.612.3M357.8M29.12×2.85%
GPT-5.3-codex (Codex)63M
Other Codex (5.2 / 5.3-spark / 5.1-codex-max / 5.1-mini)49M

Monthly token trend

MonthClaude spentClaude cacheCodex tokensMain modelNotes
2026-0138.6Mgpt-5.1-codex familyCodex starts
2026-02251Mgpt-5.2 / 5.3Warm-up
2026-03(Claude peak)(heavy cache)702Mgpt-5.3-codex / 5.4Explosion: Claude joins full-time
2026-04(main driver)(main driver)6.16Bgpt-5.4 / 5.5Two-engine maturity
2026-05(consolidation)(replay)4.41Bgpt-5.5Per-thread efficiency rises

Model migration notes

  • 2026-03: Early Codex models gpt-5.1-codex-max / 5.2-codex / 5.3-codex-spark shrink → gpt-5.4 takes over
  • 2026-04: gpt-5.5 launches and immediately carries the load (1,108 threads, 52.9% of all Codex threads)
  • Claude side: Opus 4.6 is the consistent main driver from March. From April, Opus 4.7 (1M context) joins for long-context scenarios (spent 83M / cache 1.82B), specializing rather than replacing 4.6.
  • 2026-04 → 05: Codex total tokens ↓ 28% but thread count stable — cache fluency improved, per-thread average dropped.

Unit input/output (reference only — not a KPI)

  • 100K Claude tokens per commit (429.9M / 4,292, including planning / review / implementation)
  • 502 Claude tokens per line of code churn

Reminder: AI output also includes massive high-value labor not directly converted to commits (architecture review / plan iteration / skill refactor / data cleaning / pair-debugging). Treating "X tokens per commit" as KPI is anti-incentive.

💰 Output × input

Same-period GitHub output

  • 365-day total contributions: 3,706 (commits 3,467 + PRs 114 + issues 38 + reviews 9) · repos owned: 85
  • Highest single day: 2026-03-18 (12,301 Claude messages, corresponding multi-repo GitHub burst)

Top repos (public data, original names retained)

Repolanguage365-day commitsstars
antigravity-workspace-templatePython1691,246
awesome-architectureVue37600
Readme.skill(md)24124
OpenCMOPython34984 ★
Agent_View_Controller-AVCJavaScript2546 ★
easy-claude-code(md)2445 ★
vibe-forgeRust3814 ★
clawdbot-webchat-liteTypeScript1112 ★
agent-mailer (collab)Python3811 ★
ReadYourUsersTypeScript379 ★
Vibe_coding_guide (collab)69 ★
opencrabPython2,5261 ★
Other 12 public reposvarious~2801-12 ★ each

Total stars (Top 25 cumulative): ~2,217 ★

Primary languages (GitHub primary lang + local git weight)

Python · TypeScript · Rust · Vue · HTML · JavaScript · Jupyter Notebook · Markdown (7 main + multiple secondary)

📊 Data sources & privacy commitment

  • 100% local data: ~/.claude/* (stats-cache + history + 77 projects + 11 plans + 6 skills + settings) + ~/.codex/* (state_5.sqlite read-only + history + 11 skills + 4 automations) + ~/Library/Application Support/Cursor/* (7 workspaces' ItemTable) + ~/.gemini/antigravity/brain/* (105 tasks' metadata + markdown) + 13 local git repos' git log + GitHub via gh api graphql
  • Kiro / Trae not installed; gracefully skipped (v2.5 supports them but no data on this machine)
  • Conversation bodies are read only for keyword / collaboration-style analysis — raw text never appears in the report. Antigravity reads only metadata summaries and markdown headings — no screenshots, no browser cache, no pbtxt annotations.
  • Work project names anonymized to "Project A/B/..."; public GitHub repos retain original names (public data is indexable anyway); API keys / tokens / emails scrubbed by regex.
  • Report auto-generated by Claude Code / Codex / Cursor / Antigravity local data via Readme.skill v2.5.1; fully reproducible.
  • Generated at: 2026-05-28T(local)