MyStatus Stats

The AI Stats — Token breakdowns

Before any work starts, an assistant is handed a fixed block of standing context: its own instructions, this project’s conventions, the description of every installed skill, the signature of every available tool. None of it is the work. All of it is re-read on the next turn, and the one after that.

• Tool definitionscounted80,84149%

  The full call signature of every tool loaded up front — 77 of them, 323,362 characters. One of them, Artifact, is 57,902 characters on its own. Until September 24 this row said it could not be read and carried August 30’s 14,500 forward, because tool schemas arrived with the request and were never written into the session record. They are now: the session record carries a snapshot of every tool it was handed, so the row is read like every other one here. A cloud session’s tool set is not the desktop’s — it carries the remote-trigger, remote-device and memory tools and a Projects tool — which is why this row jumped from 33,953 on September 30. On October 8 the same 46 tools came back by name for the third reading running. By October 12 the platform had changed the set it hands a cloud session: 32 tools added — display widgets, search and conversation tools, task, web and repository tools — and 5 removed, 80,833 characters more, about 20,000 tokens. None of the 32 is ours, so this row is now mostly the platform’s, and that is why the platform’s own line below fell on that reading. On October 15 four scheduled-task tools the October 12 set had dropped were back, and three of the platform’s own descriptions were longer — 21,597 characters more, about 5,400 tokens, all of it the platform’s.

• The assistant’s own system promptby subtraction56,07934%

  Anthropic’s standing instructions. Not counted directly: this is what is left after everything else is subtracted, so it also carries the error in every estimate above it. Its share fell from 54% to 40% on September 24, and almost none of that was the platform changing: the tool definitions above became readable that day and turned out to be about 19,000 tokens larger than the figure carried since August 30, all of which had been counted here. Thirteen readings now — 52%, 56%, 55%, 57%, 55%, 54%, then 40% twice on the new method, then 41% on September 30 from a cloud seat, 41% again on October 3 from the same kind of seat, 46% on October 8 from the same kind again — up five points, because the file we write ourselves shrank by far more than this line did — and 35% on October 12, down eleven. That fall is not the platform giving ground. Between the two readings it changed the tools it hands a cloud session, 46 to 73, and tool definitions are counted on their own row above, so about 20,000 tokens of the platform’s moved out of this line and into that one; its own instructions got shorter at the same time. Then 34% on October 15, down one: this line barely moved, 55,962 to 56,079, while the platform’s returning tools again landed on the counted row. The session record also carries these instructions in full, the second copy of the skill descriptions, and on the cloud seats a memory block of about 9,000 tokens; none is counted yet, on purpose, so each reading stays comparable with the one before. Read this line as a ceiling on the part nobody here can edit, and a generous one.

• Project instructionscounted19,15911%

  This repo’s CLAUDE.md. Conventions, invariants, and every gotcha learned the hard way — the file that lets a session pick up where the last one stopped. 76,636 characters on October 15, up from 75,704 on October 12 and 67,514 on October 8, and still under half the 158,415 of October 3: on October 4 the file was put on a diet, every rule kept to its statement and the story behind each moved to a history file a session reads only when it needs the reason. Ten readings up without once going down, the eleventh cut it by more than half, and the twelfth and thirteenth found it growing again. Still the largest line we write ourselves.

• Skill descriptionsestimated4,7672%

  34 installed skills on this reading, the same 34 as October 12 to the character, against 31 on October 8 and 86 on the desktop seat before the move to the cloud — a cloud session is not given the same install. Each description is presented twice, in two different formats. One copy is read exactly — 19,067 characters — and the second is estimated at the same size, which is why this row is still marked estimated. That estimate is known to be low: the second copy has never been counted, and sits in the platform’s line below.

• Connector instructionscounted1,4800%

  How to use each connected service. Four blocks on this reading, 5,920 characters, the blocks added up. No database connector this time — since October 3 those stay off unless a task needs the database — where October 8 had one and October 3 two. One other service the October 8 reading carried is no longer attached, and the MyStatus connector’s block is a little longer again, 959 characters, for its tool that lists the tasks a key reaches.

• Reserved tool namescounted1,1560%

  129 tools listed by name only, loaded on demand. None of them is a database connector’s; on October 8, 21 were, and without them that reading would have been 1,183. On October 15 four scheduled-task tools left this list because they are now loaded up front, and one MyStatus tool joined it. Still among the cheapest things in the list. The names as injected, one per line with a newline between each: 4,622 characters.

• Subagent listingcounted6330%

  The six kinds of helper session that can be dispatched, each with a line saying what it is for. Counted separately rather than left inside the remainder below — a residual should hold only what genuinely cannot be read. The same six, 2,532 characters, identical to the character on six readings running, across a change of platform — the one control that has never moved. The six lines add up to 2,527 on their own and 2,532 with a newline between each; every reading since September 28 has used the second count.

The total is measured; the breakdown is not, evenly. The 164,115 is exact, read out of a session's own transcript. The parts above are good to about ten or fifteen percent, so trust their order rather than their digits.

A dated snapshot, not a live counter: the transcripts it reads sit on one laptop and never enter the repository, so one person refreshes this by hand. Last on Oct-15-26.

Thirteen re-counts now. Six moved the share by a point or two, the seventh moved the method, the eighth held, the ninth moved the seat, the tenth held on that seat, the eleventh went up because our own file went down, the twelfth went down because the platform’s tools moved into a row we count, and the thirteenth held. Every part above was measured again on Oct-15-26. On August 30 raising the total alone would have pushed the share we do not control from 56% to 65%, and counting properly sent it down to 52%. On September 5 the shortcut said 57% and the honest count said 56% — very nearly the same, and up rather than down. On September 7 the shortcut said 57% again and the honest count said 55%, down by one. On September 13 the shortcut said 59% and the honest count said 57%, up by two. On September 18 the shortcut said 60% and the honest count said 55%, down by two. On September 21 the shortcut said 56% and the honest count said 54%, down by one. On September 24 the shortcut said 56%, the count done the old way said 53%, and the count published said 40% — because a row this page had carried forward for four readings as unreadable could be read that day, and was more than twice the figure it had been carrying. That drop was the method changing, not the platform shrinking, and it is not comparable with the six before it. On September 28, on the same method, the shortcut said 42% and the count said 40% — the first reading since the change that can be compared with the one before it, and it barely moved. On September 30 the session refreshing this page was a Claude Code seat, which opens at about 113,000 tokens against the desktop's 155,000, so the count was taken from a cloud session instead, which has its own tools, skills and connectors: the shortcut said 46% and the count said 41%, and the parts moved by about as much as the seat did. That reading does not compare with the eight before it, and the Claude Code reading is on the chart beside it rather than standing in for it. On October 3 the count was taken from the same kind of cloud session as September 30's, so for the first time since September 28 a reading compares with the one before it: the shortcut said 43% and the count said 41%. That session had two database connectors attached that are now off by default, about 1,100 tokens of the total. On October 8, from the same kind of session again, the total fell for the first time in the series — 174,088 to 149,621 — because the file we maintain had been cut by more than half four days earlier: the shortcut said 32%, the platform down to a third, and the count said 46%, up five points. The platform's own figure fell too; it was a bigger share of a smaller floor. On October 12, from the same kind of session, the total rose to 158,365 and the count said 35%, down eleven, while the shortcut said 49%. That is not the platform's part shrinking. It moved: the platform now hands a cloud session 73 tools where it handed 46, about 20,000 tokens more, and tool definitions are a row counted above, so the leftover got smaller while the platform's own instructions in it got shorter too. On October 15, from the same kind of session again, the total rose to 164,115 and the count said 34%, down one, while the shortcut said 37%: four scheduled-task tools came back into the set the platform hands a cloud session, about 3,900 tokens on the counted tool row, and the leftover barely moved. All thirteen are published. A method that only ever rescued us from an unflattering number would be worth suspecting, and the reason the share kept moving before the method changed is that the part we wrote ourselves grew faster than anything else on the list: the file we maintain went up at the first ten of these readings, down at the eleventh and up again at the twelfth and thirteenth, and the share moved the other way each time.

How far to trust the reading. The parts above are read out of the session's own transcript, character for character, rather than eyeballed — and the check on that is a row whose set did not change between two readings, which should come back at the same figure. On the reading dated above the seat was the same kind as the one before, and two rows qualified: the helper-session list, the same six lines, came back identical to the character for the sixth reading running, and the skill descriptions, the same 34 skills, came back identical to the character too. The tool definitions did not qualify, because the platform changed the set again — four tools back in — and the row says so. And before the count was pointed at this reading it was run over the previous one's transcript copy, where it reproduced every figure published on October 12 to the character. That says the reading works. Every other row moved because its set did — tools returned by the platform, our own file and our connector's block a little longer, one reserved name traded for four — and each row says which. One correction made on September 28: the helper-session list's five-character move on September 24 was two ways of counting the same six lines, added up and with a newline between each, not the platform changing. How each row is counted is now written down beside the figures, so the next reading counts the same way. The platform's own line is the least firm one. It is what is left after everything else is subtracted, and three things that are now readable still sit inside it uncounted — the second copy of the skill descriptions, the platform's instructions themselves, and on the cloud readings a memory block of about 9,000 tokens — so it is a ceiling, not a measurement.

Measured task by task, not estimated. Every task closes with its own turn count and scaffolding figure on the MyStatus Work Log, so the figure above is checkable against its source rather than asserted here. There are 385 so far, which is why it reads as a range. The list below carries 386; the remaining one closed inside a single second, so it has no floor to average.

• Two processes, one stampclaude-opus-5-56,370,977

  69 turns at 92,333 each. The entry it came from

• Two processes, one stamp, read on live (a followup)claude-opus-5-52,123,659

  23 turns at 92,333 each. The entry it came from

• Read by right leaves a rowclaude-opus-5-511,918,080

  128 turns at 93,110 each. The entry it came from

• The card footnote is a named partclaude-opus-5-514,642,355

  91 turns at 160,905 each. The entry it came from

• Read by right leaves a row, walked on live (a followup)claude-opus-5-53,072,630

  33 turns at 93,110 each. The entry it came from

• The seat's own card opens: the cloud block runs as printed (a followup)claude-opus-5-56,114,390

  38 turns at 160,905 each. The entry it came from

• The rehearsal holds the candidate out of the chainclaude-opus-5-53,679,280

  40 turns at 91,982 each. The entry it came from

• Every printed card opensclaude-opus-5-55,518,920

  60 turns at 91,982 each. The entry it came from

• Every tool error a seat meets is countedclaude-opus-5-57,431,414

  82 turns at 90,627 each. The entry it came from

• Admins manage, leaders readclaude-opus-5-519,882,472

  218 turns at 91,204 each. The entry it came from

• The watchersclaude-opus-5-522,490,219

  223 turns at 100,853 each. The entry it came from

• The names rosterclaude-opus-5-516,370,152

  164 turns at 99,818 each. The entry it came from

The 12 most recent of 386 finished tasks. The figures above are computed over all of them; only this list is shortened. Every one of them is on the Work Log, in full and in order.

The model is a receipt, not an explanation. Naming it removes one thing a reader would otherwise have to guess at when two rows here disagree — different models use very different amounts — and that is the whole of what it claims. It does not explain the variance: these tasks also differ in how big they truly were and in how much of each one went on debugging, and no figure on a row separates those. It records what ran, not a choice anyone made; a session’s model is set outside the conversation. Rows that closed before the entries carried a model do not have one here either.

Why publish the overhead at all

Hours on the record

Two things a clock can answer: how long the work was open, and whose time it was. Every session is a sequence of timestamped records, so the wall clock splits in two — time the assistant was producing, and time it was stopped, waiting on someone to read and decide.

Two questions, two denominators. The share is of SESSION time: a minute with one session thinking and another coding is honestly one of each, so the parts sum to 300.3 hours. The headline is the CLOCK, where that same minute is one minute — 243.5. Neither is derived from the other's denominator, and the 56.8 hours between them is time two sessions were live at once.

This is a floor, and it is a floor three times over. It counts only ATTENDED review — the stretches where the work was demonstrably stopped, either just before someone typed or inside a question that suspends until it is answered. Reading along while the assistant works does not count. Thinking away from the keyboard does not count. And any gap over 45 minutes is dropped rather than billed as deliberation, which discards 406.1 hours — more than it keeps. Every one of those choices makes the figure smaller.

These are a floor on the effort, too. Verifying a build on a phone, reading away from the keyboard, and thinking between sessions all happen where no transcript can see them, and none of it is counted here. 143 sessions, last read Aug-30-26.

Estimated by work type

A later refinement of the analysis above, begun Aug-23-26 to give a clearer breakdown of what the hours were spent on. It reads a different instrument — the public work log, where a task posts a row when it opens and a row when it closes — so it starts later and it sees less.

• Coding481 sessions293.0 hrs

  Averaging 37 min. 90% of them have both ends posted; the rest ran out of beats and were closed by the silence rule. MyStatus Work Log

• Business372 sessions330.7 hrs

  Averaging 53 min. 45% of them have both ends posted; the rest ran out of beats and were closed by the silence rule. Biz-zy Work for MyStatus

• Both at oncecounted once− 106.4 hrs

  Time when a coding task and a business session were open together. Real work in both columns, one hour on the clock.

What this breakdown misses, over its own dates. From Aug-23-26 onward the transcripts see 95.2 hours on the clock, and the lines above attribute 517.3 of them to a kind of work. The rest is real work that posted no row — a scheduled task, a one-question session, or the stretch between one task closing and the next opening. It is counted in the headline above and not here.

These are a floor on the effort. Verifying a build on a phone, reading, and thinking between sessions all happen off-stream, and none of it is counted here.

A session ends where the record says it ends: on a posted “Closed” or “Stopped”, or — when it simply went quiet — 30 minutes after its last beat, the same end a reconstructed stop is given. Once a stream has gone quiet long enough that the sign at the top of this page stops believing it, the session is over rather than counted through: a laptop closed overnight costs half an hour, not eight.

The rate

Want this for your own project — the building, or the receipts that prove it? Head over to Omega Ortega to chat more about that.

The team

The split, as of today: Claude Fable 5 takes architecture, refactors, and strategy; Claude Opus 5 gets handed straight coding tasks. Everything either of them produces goes past one person who vets the logic, questions the thinking, and owns every decision that ships — which is the part that makes the first number a 1 rather than a 0.

The build

Counted from the repository at commit b4c8c9b on Oct-18-26, 18:58 UTC, from the claude branch. These are generated by hand rather than at deploy time, and they are always exactly one commit behind — a commit cannot contain its own count.

The biz

How fast we learn

Read from the public Lessons task, which carries clean copies of every lesson our 149 seats posted and every row that used one, on Oct-11-26, 02:37 UTC, at 14555e2. A run by hand, not a read per request, so this is as of that moment.

The service

We’re showing these numbers because we’re a new biz. We want to be upfront that this is not an at scale project yet and that’s fine because we’re very early in our experiment.

As of Oct-19-26, 13:39 UTC. Ops tasks is the one number here you can recount yourself: every task in it is listed at /browse/ops, and what they actually do is written down.

Construction downtime

What is not here

No reader counts, no page views, no “most popular tracker”. Reading a MyStatus link needs no account and leaves no trace a reader would have to think about, and a leaderboard would quietly change what this product is for. There is also no “features planned” number: the roadmap lives in prose, and a hand-typed figure that drifts every week is worse than no figure at all.

Inspiration

This section exists to honor Cluetrain Manifesto’s reminders to be human, drop the corporate act and acknowledge there are no secrets. That’s why I’m building this in public and completely documenting how I dogfood the tool to build the tool.

It also honors John Wilker, my friend and old biz partner who loves Cluetrain as much as I do. He once shared how he has a dashboard of all his writing stats. It helps him write awesome booksyou should go buy! I made my own stats dashboard, which I now share with you.

Stuff only math nerds care about