Turning scattered information and unfinished ideas into systems I can keep using.

A unified record of local collaboration, tasks, and workflows across Claude Code and Codex. Translation, narration and social scheduling run through separate script pipelines and are not part of this conversation record.
Start with six examples: reconnecting knowledge, reading across languages, listening to articles, reading along a thread, checking model choices, and exploring an event archive. Each has an entry point and a clear scope.

Start with an example

Ongoing systems and completed experiments are labelled separately.
Content website

Read across languages

Articles, related concepts and language editions in one place.

The problem Scattered content makes it hard to keep following a question.

Try this

Open an article, switch languages, then follow a related link.

Start reading →

Scope:Translations and content are revised; language editions are not separate outcomes.

Knowledge system

Reconnect earlier knowledge

Follow links between concepts and sources to find, check and extend an idea.

The problem Saved information can be difficult to find again.

Try this

Search the Wiki for a topic such as human–AI collaboration, then open a concept and its sources.

Open Wiki →

Scope:Search covers collected, public material, not an exhaustive knowledge base.

Content pipeline

Listen to articles

Articles narrated with a cloned voice, published as a podcast and on article pages.

The problem Long reads do not always fit into a sitting.

Try this

Subscribe to the podcast on the subscribe page, or press play on an article that has audio.

Subscribe to audio →

Scope:The voice is an AI clone, not a studio recording; not every article has an audio version.

Curation layer

Read along a thread

Articles regrouped into series and core ideas so one question can be followed end to end.

The problem Once there are many articles, a flat list hides how they relate.

Try this

Pick a series, or start from a core idea that runs through several articles.

Browse series →

Scope:The grouping is editorial, not algorithmic; series are still being added.

Data dashboard

Check model choices

The models in use and under evaluation, kept as a selection map that is re-checked over time.

The problem Models and prices move fast, and choices drift toward habit.

Try this

Look up the current suggestion for writing, research or coding, and the date of the last check.

Open the model map →

Scope:A snapshot from the last check; it does not replace testing on your own task.

Completed · 2026

Explore a pilgrimage archive

An archive of an experiment in pilgrimage activity and ESG records.

The problem Event records can become scattered once an event ends.

Try this

Explore the retained material to understand the setting of the experiment.

View the archive →

Scope:A historical archive, not a live service for a current event.

Decision records · Avoiding dependence on a single git host
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.
Decision records · De-identified does not mean publishable
De-identified does not mean publishable
ContextIn July 2026, seventeen de-identified daily digests were judged safe to publish and were withdrawn the same day they went live.
ConsiderationsTwo options: keep the flow of de-identifying first and then assessing publication, or treat any material with a privacy concern as private by default.
ApproachThe latter. De-identification only answers whether something is safe to keep in the repository; it does not unlock publication. The rule applies to every public channel on the site, not only the wiki.
ResultThe rule is written down as a decision record and a project instruction. Similar material now stays private by default instead of going through a de-identify-then-assess loop.
Decision records · Check deployment lineage before shipping
Check deployment lineage before shipping
ContextIn September 2026, a batch of articles deployed from a branch was never merged back to main. A few days later a deploy from main overwrote them.
ConsiderationsRely on people remembering to merge first, or put the check inside the deploy script so it blocks on its own.
ApproachThe script now runs two checks: the last successful deploy must be an ancestor of this one, and this version must be reachable from main. If either fails, nothing is deployed; a deliberate rollback needs separate authorisation and a logged reason.
ResultDeploying straight from a branch or worktree is now blocked, with tests for the rule. Live content being overwritten by an older version has not happened since.
Decision records · Retiring the fully automatic self-optimisation engine
Retiring the fully automatic self-optimisation engine
ContextAn experimental engine let the site propose and test its own changes. It stalled quietly within seven weeks; its experiment log stopped in late March and nobody noticed.
ConsiderationsFix it and keep running, or accept that a personal site gives too sparse a reward signal to support fully automatic iteration.
ApproachRetired in May 2026 and archived as a whole. Replaced by a division of labour between several AI tools and me: each researches and proposes, and a person still decides.
Resultllms.txt, MCP and structured data were downgraded to basic hygiene rather than a moat; the workflow now centres on documents and decision records.

Work records and data

These counts describe activity and records, not quality or labour saved.Aggregate snapshot: 2026-11-21 07:41 UTC+08:00AI source records since 2026-02-10 (285 days); Git main since 2026-03-19 (248 days)

Data period: 2026-03-19 to 2026-11-21

Days with non-automated commits
230 / 248

Dates with non-scheduled commits, not full days worked.

Handoffs: cross-session briefs (total)
492

Documents intended to continue work across conversations, not a handoff success rate.

Decision records
48

Decision records preserving the reasoning, included in governance documents.

Cumulative net line changes1158K lines · accumulated from version history
0K100K200K300K400K500K600K700K800K900K1000K1100K030405060708091011

Includes code, content and documents, excluding scheduled commits. This is neither the current codebase size nor a measure of value.

Non-automated commit calendarLess More Platform activity
Jan 2026FebMarAprMayJunJulAugSepOctNov

Grey means covered with no non-scheduled commits; stripes mean not covered, and faded cells are future dates. Select a cell for details.

2026-11-21 · 3 commits

Git main has commits on 248 of 248 days; this calendar shows non-automated commits, so 18 automation-only days are blank.

AI foreground time on Mac

Tracked by Timing · Mac only · Taipei time

Loading

Total AI time
Pending
Past 7 complete days (excluding today)
Pending
Today
Pending
Source month
Pending

Apps
    Pending
Websites
    Pending
Devices
    Pending
Mac only; iPhone and other devices are outside this metric.

Bars rank hours; each chart uses its own maximum. These are not shares of total time. Do not add apps, websites and devices together.

Counts specified AI apps and websites in the Mac foreground, including reading, typing and waiting; overlapping intervals count once. Excludes phones, background execution, unidentified terminal use, general IDE use and whole AI-browser sessions. This measures identifiable AI use, not every conversation or AI compute time.

More time may reflect learning, production or rework; hours alone do not establish efficiency.

Version activity and workflow details

Overview

Sources include git plus local Claude Code and Codex usage records
Non-scheduled commits
7,149
Deployment-related changes (proxy)
1070
File-history changes, not successful deployments; 67 PR merges recorded separately.
Code & content changes
+1460K / −301K
lines (added / deleted)
Days with non-automated commits
230 / 248
230 non-automated, 18 automation-only, 0 with no commit across 248 Git days
Consecutive days with version activity
248d
Includes automation · strict human–AI activity 130d

Workflow records

Cross-window AI agent collaboration, organized into records that can continue, hand off, and be tracked
Worklogs: work journal (total)
1036
Handoffs: cross-session briefs (total)
492
Governance docs (incl. ADRs · total)
738
48 of which are ADR decision records
AI conversations and token definitions

Data period: 2026-10-22 07:38 UTC+08:00 to 2026-11-21 07:38 UTC+08:00

These token counts use different definitions; do not compare efficiency, add them or infer costs.

Claude Code root sessions (rolling 30 days)
543 sessions
26,169 messages · Opus 5.5
plus 153 subagent runs, reported separately
Codex root tasks (rolling 30 days)
108 sessions
1,631 messages · GPT-6.1 Sol
plus 46 subagent runs, reported separately
Claude I/O tokens (rolling 30 days)
16.8M

Deduplicated input + output; cache tokens excluded.

Codex token delta (rolling 30 days)
975.1M

In-window increments of cumulative usage snapshots.

Peak collaboration hour
09:00
Claude + Codex root messages
Tests and deployment process data

Verifiable Quality Signals

Scopes differ; missing data does not mean zero.
Test coverage
Python 85.9% · Node 51.4%

Measured: 2026-11-21 07:34 UTC+08:00

Deploy smoke-test compliance
87%
Cross-project impact-tag compliance
100%
Contract-test SLA adherence
91%
Reverts / rollbacks
24
Test files
136
Manual deployment process failure rate
16.4%

Since 2026-06-21 16:14 UTC+08:00 · 654 manual deployment records. Deployment command or post-deploy check failures; later hotfix attribution is not included.

Recorded incidents
18

Historical records since 2026-06-22 10:37 UTC+08:00; missing recovery times do not establish that incidents remain unresolved.

A separate 217 scheduled deployment records include 19 process failures; they are excluded from the manual sample.

Delivery measurement limits

No individual rating
Incident recovery time
not measured
Change lead time
not measured

This process failure rate includes deployment-command and post-deploy check failures, but does not attribute later hotfixes. It differs from DORA change fail rate. DORA uses five delivery metrics; this site does not derive ratings from proxies.

DORA

Five-Dimension Capability Evidence

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
Deployment process records exist; successful deployment frequency, recovery and lead time require separate evidence.
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
These are mainly test and process signals; high coverage does not establish test quality.
Influence | outward reachSelf-reported
No external-reach signal; left honestly blank.
Data observations

Data observations

Consistency

Version activity spans 248 consecutive days including scheduled jobs; human commits or root AI conversations span 130 consecutive days. This is not service availability.

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

Conversation counts cover local Claude Code and Codex only; screen time comes separately from Timing. Token definitions differ and must not be compared or added. Hours, lines and commits do not directly measure efficiency.

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.
  • Direct records, proxies and estimates are labelled separately. Unknown is not zero; no ability score or labour saving is inferred.
  • 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.
ToolWhich AI model is strongest, and best value, right now?A model comparison with source dates: general intelligence, coding, long-horizon agentic, and the price-performance frontier, all on one page.See the model frontier →