Turning some long-shelved ideas into systems that actually run.

A unified record of local collaboration, tasks, and workflows across Claude Code and Codex.
This is a record of how I've been working with AI lately: some of it is websites, some knowledge systems, some data wrangling and automation. They aren't necessarily complete, but each has gone from an idea to something usable and maintainable.
Paul Kuo（paulkuo.tw）
ToolWhich AI model is strongest, and best value, right now?A weekly-refreshed frontier snapshot: general intelligence, coding, long-horizon agentic, and the price-performance frontier, all on one page.See the model frontier →

Projects I've worked on

A few examples that are live or were once running.
paulkuo.tw multilingual content site
My content site, with versions in several languages, tying together caching, the Dedao brain API, a Worker API, dynamic share images, and login. It mainly hosts articles, knowledge notes, and project records.
paulkuo.tw ↗
LLM Wiki personal knowledge system
Turning notes and material from different sources into a knowledge system you can search, ask, and visualize. The point isn't to show off a graph, but to let the material be re-linked, found, and extended.
paulkuo.tw/wiki ↗
Baishatun Mazu ESG pilgrimage record platform
An activity-record platform set around the Baishatun pilgrimage, combining multilingual content, LINE check-ins, route data, and carbon- and kindness-footprint records. The project wrapped up in 2026; the site stays online as a record of the experiment.
mazu.today ↗

Decision records

One decision left behind during the work.
Avoiding dependence on a single git host
ContextThe main git-host account was once suspended, which made me realize that if version history, deploys, and issue records all depend on one platform, that's a single point of risk.
ConsiderationsAt the time I weighed three options: keep using one platform, move everything to another, or set up several mirrors so different platforms back each other up.
ApproachI chose to set up mirrors across platforms. If any one platform has a problem, the others still keep the version history and the ability to update; deploys moved to wrangler, not tied to any specific git host.
ResultAfter the account was restored, I brought it back as a third mirror. Since then, no single git host is the only dependency, and issues are just one part of the work record rather than the whole truth.

AI collaboration — raw records

Some of the raw records behind the projects and workflow above.Last updated: 2026-10-14 07:50AI source records since 2026-02-04 (253 days); Git main since 2026-03-19 (210 days)

Overview

Sources include git plus local Claude Code and Codex usage records
Commits
5,266
Deploys
791+
plus 33 PR merges
Code & content changes
+1048K / −213K
lines (added / deleted)
Days with non-automated commits
191 / 210
191 non-automated, 18 automation-only, 1 with no commit across 210 Git days
Collaboration system uptime
209d
Includes automation · strict human–AI activity 123d
Claude Code root sessions (rolling 30 days)
635 sessions
26,462 messages · claude-opus-5-5
plus 498 subagent runs, reported separately
Codex root tasks (rolling 30 days)
157 sessions
3,268 messages · gpt-5.6-luna
plus 61 subagent runs, reported separately
Claude I/O tokens (rolling 30 days)
21.5M
Codex token delta (rolling 30 days)
1395.9M
Peak collaboration hour
22:00
Claude + Codex root messages
Codebase Growth836K lines · from zero
0K100K200K300K400K500K600K700K800K0304050607080910
Non-automated commit calendarLess More Platform activity
Dec 2025Jan 2026FebMarAprMayJunJulAugSepOct
Git main has commits on 209 of 210 days; this calendar shows non-automated commits, so 18 automation-only days are blank.

Workflow records

Cross-window AI agent collaboration, organized into records that can continue, hand off, and be tracked
Worklogs: work journal (total)
953
Handoffs: cross-session briefs (total)
379
Governance docs (incl. ADRs · total)
414
47 of which are ADR decision records

AI time on screen

Tracked by Timing, deduplicated across Mac and iPhone

Loading

Total AI time
Pending
Last 7 full days
Pending
Today
Pending
This month
Pending
Apps
    Websites
      Devices
        When both devices are in use at the same time, each counts it, so the device totals add up to more than the overall total.

        Screen time counts only the time an AI tool is in the foreground, and overlapping time across devices is counted once. Time AI spends running in the background is listed separately in the estimate card and is not mixed into these figures.

        Verifiable Quality Signals

        Raw evidence layer; not-yet-accumulated is shown as “not measured”, never 0
        Test coverage (Python / Node)
        80.9 / 42.4%
        Deploy smoke-test compliance
        93%
        Cross-project impact-tag compliance
        100%
        Contract-test SLA adherence
        93%
        Reverts / rollbacks
        21
        Test files
        78
        Change failure rate
        13%
        Incidents / MTTR
        15

        Delivery Signals (DORA, directional)

        Solo personal site, manual deploy: directional only, no Elite badge
        Deploy frequency
        4.14
        daily~on-demand
        Change failure rate
        13%
        Incident recovery (MTTR)
        not measured
        Change lead time
        not measured

        DORA is a team software-delivery scale; solo direct-to-main inflates frequency/lead-time semantics, and tiering needs all four metrics together. Treated here as a directional reference only: no tier, no personal badge.

        Five-Dimension Capability Evidence

        Qualitative read + evidence level; no composite score, no grading
        Command | steering the AI collaboration system (direction / governance / orchestration)Artifact-evidenced
        Ample governance / orchestration artifacts; command quality is self-reported.
        Delivery | speed and stability of shippingPartial
        Deploy frequency and CFR now have signals; MTTR and lead time still show gaps when evidence is absent.
        Leverage | amplifying output with AIArtifact-evidenced
        High output with Claude and Codex usage shown separately; a multiplier would need an unmeasurable no-AI same-person baseline.
        Quality | reliability of outputAuto-fetched
        The hardest dimension, but mostly process signals (high coverage != good tests).
        Influence | outward reachSelf-reported
        No external-reach signal; left honestly blank.

        Data observations

        Consistency

        The collaboration system has run for 209 consecutive days, including scheduled automation. The stricter record—at least one human commit, Claude Code root prompt, or Codex root user message—has continued for 123 days.

        Collaboration records

        AI conversations break easily, so I keep handoffs, ADRs, and work journals on the side. These records let separate sessions pick up where the last one left off, and let a later me see why something was done the way it was.

        Data limits

        This covers aggregate local Claude Code and Codex records, not Chat, Cowork, or other AI tools. Their token definitions differ, so they are shown separately and never added together; deploys and lines changed remain rough references.

        How the data is compiled

        • Sources include the git mainline plus raw local Claude Code and Codex session records.
        • Commit counts come from a single git mainline. GitHub, GitLab, and Codeberg are mirrors of the same history and aren't counted twice.
        • Commits, active days, heatmap, and changed lines exclude scheduler commits. System uptime includes scheduled commits; strict human–AI activity uses human commits plus Claude Code root prompts and Codex root user messages.
        • AI figures use a true rolling 30-day window. Root sessions feed the main figures; subagents are reported separately to prevent double counting.
        • Claude tokens are deduplicated assistant-message input + output; Codex tokens are window deltas from cumulative task snapshots. The two definitions are not added together.
        • Repo platform activity is traceable from 2026-02-20 via PR and issue records; because the git mainline history was reset in mid-March, commit details before 3/18 aren't included in the current counts.
        • This page makes no scores or estimates; it only shows data that can be checked against the raw records.
        • Workflow artifacts recursively include subdirectories and Codex reviews while excluding backups, archives, drafts, and prompt templates. ADRs are included in governance totals, not counted twice.