# What changed in: How much of this did I write? The numbers behind twenty articles in four weeks, and what the input actually was, sgit.ai > The changes to the article "How much of this did I write? The numbers behind twenty articles in four weeks, and what the input actually was" between two published versions, paragraph by paragraph, with insertions and deletions marked. *Source: · site v0.6.75 · this file is generated from the same content as the page, so the two cannot drift. Every page on this site has a `.md` twin; internal links below point at them.* --- [Home](../../index.md) / [Articles](../index.md) / [How much of this did I write? The numbers behind twenty articles in four weeks, and what the input actually was](../how-much-of-this-did-i-write.md) / Changes # What changed in: How much of this did I write? The numbers behind twenty articles in four weeks, and what the input actually was The article [How much of this did I write? The numbers behind twenty articles in four weeks, and what the input actually was](../how-much-of-this-did-i-write.md), compared paragraph by paragraph between two versions. From: d234ea8cf, 2026-10-05, site v0.6.66: the map redrawn, the workflow named, the changes shown. To: the working tree, the version being built now. 14 paragraphs added, 0 removed, 16 changed in place, 56 unchanged. About 1,492 words added and 0 removed. Insertions are marked like this, deletions like this; unchanged runs are folded to one line. Figures appear as their file names. Generated by admin/build/article_diff.py from the git history. 1 unchanged paragraph Summary: A friend who received a reply from one of my agents said the amount of text I manage to produce is baffling, and the honest answer deserved numbers rather than a shrug. So this article measures the session that wrote the last twenty articles on this site: every word I typed or spoke, every word that came back, how many times each piece went round, what the corrections were, and what the memos were standing on. The picture is not the one people assume, in either direction. I did not type the articles, and the model did not write them from a prompt. Over four weeks I sent about 63,00065,000words,words53,000in 148 messages, 55,000 of them in thirty-fivethirty-seven voice memos, and the agent published 85,00096,000 words of articles and wrote another 74,00076,000 to me about them, through 122123 releases, 135 web searches and 2226 research agents. No article came from a one-line prompt; the shortest brief was a single memo of 1,943 words, the longest ran to twenty-ninetwenty-one messages. SevenEight articles are accounted for by hand, memo by memo and correction by correction.correction, and every number is a row in a published vault. And the input that matters most is not in the session at all: the last article quotes thirty-three pieces of my earlier writing, from a 2010 open source tool to briefs written with other agents this summer, which the agent found because they were published. The corrections I make are rarely to hallucinations. They are to briefs that needed to be better, because a model that can go in any direction needs someone with a direction. That is why the people who have one are not out of a job. They are the input. 1 unchanged paragraph > **Where this comes from, and what is behind it.** A WhatsApp exchange on 5 October 2026, in which a friend who had just received an answer from one of my agents said the amount of text I produce is baffling, and my reply that I create the briefs and review the drafts, and that the article in question leaned, in a Wardley-evolution sort of way, on content I have been publishing for years. Then a voice memo asking for the thing I always try to explain to be measured rather than asserted. The data is the transcript of the Claude Code session that has run this site since 11 August, parsed for who said what and when, joined to the site's git history. The method, its limits and every number are in [the evidence vault](../../demos/vaults/how-much-evidence/index.md) behind this article, built on the fullsame day by a data scientist agent and a developer agent from a brief, with the app that reads it; a summary dataset areis[published beside[beside the article](../../articles/data/how-much-of-this-did-i-write.json). SevenEight of the twentytwenty-two articles were checked by hand, turn by turn; the rest are reported only where the numbers are exact. This article has itself been revised since it was first published, and [what changed](../../articles/diffs/how-much-of-this-did-i-write.md) is shown paragraph by paragraph, by a tool built for the purpose on the same day. 1 unchanged paragraph, under In short • **The scale.** In twenty-seven days I sent 281148 messages, 62,89665,323 words in all, of which 52,71955,473 were in thirty-fivethirty-seven long voice memos, plus fifteeneighteen uploaded files and thirty-twoforty-eight images. The agent published twentytwenty-two articles totalling 85,20996,451 words with ninety-four105 figures, wrote 74,45675,993 words to me about them, and made 122123 releases of the site. It ran 135 web searches, 190 page fetches, 1,9522,001 shell commands and 2226 research agents to do it. • **No article came from a one-line prompt.** The shortest brief any article had was one memo of 1,943 words with nothing after it; the longest ran to twenty-ninetwenty-one messages. The median article went through between two and three releases.releases, and the one that went through nine spent four of them on its title. The words-out to words-in ratio runs from about one to about four and is the least interesting number here. • **Seven**Eight articles, accounted for.** The fractal semantic graphs article took 2,395 words from me in twenty-one turns, including a 1,060-word message that began "this is still wrong", and came out at 2,447 words over five releases. The memory article took 1,099 words in five turns and came out at 4,401 in one. The code review company article took two memos of 3,312 and 1,072 words, a reader's 1,401-word email, three rounds of additions, and came out at 9,289. The corrections I make arrive within the hour and are carried in minutes: a median of thirty-eight minutes from a release to my next correction, and five minutes from the correction to the release that carries it. 1 unchanged paragraph • **The input is not the memo.** The last article quotes or links thirty-three pieces of my earlier writing. The record it drew on is a 2010 open source tool, a book, 107 posts on diniscruz.ai in twenty months, 3,025 markdown files in one team's record this year, and the twenty-six articles before it on this site, seventeen of which cite an earlier one, with sixty-five links between them. The agent is fast because the record exists and is published where a machine can read it.it, and the record is also a better thing to learn from than anything a model can find on its own. 6 unchanged paragraphs, under The question, and why it needs numbers, What went in, and what came out Across the window from 9 September to the evening of 5 October I sent 281148messages.messages,Stripped65,323ofwordssysteminnoise they total 62,896 words,all, and the distribution is the first thing worth seeing: the median message is fifteensixty-seven words long, and thirty-fivethirty-seven messages of more than four hundred words carry 52,71955,473 of the total.total, eighty-five percent. Those thirty-fivethirty-seven are the voice memos, transcribed as they came, with the ums left in. Beside them went fifteeneighteen uploaded files, mostly review packs from other agents and documents I had written elsewhere, and thirty-twoforty-eight images, mostly screenshots of something that looked wrong and infographics other models had made. Of the 148, forty-six are corrections and five are plain approvals, and fifty-two open with an approving word and then ask for something, which is the shape of the typical message: good, now change this. What came back was twentytwenty-two articles totalling 85,20996,451 words with ninety-four105 figures, each figure drawn as a web page in a shared style and rendered to an image, each article with a typed graph of its ideas beside it. To make them the agent wrote 240243 files, made 382 edits, ran 1,9522,001 shell commands, fetched 190 web pages, ran 135 web searches and launched 2226 research agents of its own, each of which read for ten to twenty minutes and reported back.back, all of them on the six days an article went out. It wrote 74,45675,993 words to me, whichaislittlealmostmoreasthanmuchIaswrotethetoarticles,it, in the form of plans, findings, and the end-of-turn summaries that say what was done and what was found. And the site went through 122123 releases, each with a version number, a log entry and a commit carrying both our names. !shot hmi-seven-articles-ledger.webp | images/ | SevenEight articles accounted for by hand.hand, including this one. Input is every word I typed or spoke to the agent about that article, from the first memo to the last correction. Output is the published body. Rounds are releases that changed the article's file. The lighter bars are material I brought that I had written before the session: a design pack of five documents, and a reader's email. SevenEight articles are accounted for by hand in the figure above, and they show the range. The fractal semantic graphs article, in September, is the one with the most back and forth: two memos and nineteen more messages, 2,395 words from me, 2,447 words out, five releases, and in the middle of it a message of 1,060 words that begins "Ok, but this is still wrong" and explains, with examples, why the folder-tree analogy the draft was using confused depth with meaning. That article is 2,447 words long because it was argued down to the right ones. The SaaS article a day later is the opposite shape: one memo of 2,797 words and then nine short messages, four of them about the title,title and two about the hero image, which went from one that said the apocalypse was optional to the one that says it will be decided by inertia, not by AI, with "the 3rd one nailed it" as the sign-off. The October articles are longer and took fewer rounds, which is partly the subjects and partly four weeks of the agent learning how I want things to read. The memory article took one memo of 1,039 words and four short messages, and came out at 4,401 words in a single release, with no corrections of substance. The code review article took one memo of 1,843 words and ten messages and came out at 4,989 words in three releases, along with a vault of twenty files in which a real codebase was read as layered graphs, which the memo asked for in one sentence. The reply-tail article took the shortest memo of the seven,eight, 451 words, and then elevenseven messages, one of which was me rewriting a paragraph myself because the draft had my experiment in the wrong voice, and three of which were about an image another model had made, rejected twice before it was right. The two from 5 October show the two things the ratio hides. The identity article took 398395 words from me in the session, which would make it look like the agent did everything, except that the memo came with a zip of five design documents, 17,503 words, that I had written that week with other agents, and the article is the story of those documents. The code review company article took the most from me, two memos of 3,312 and 1,072 words, and a reader's email of 1,401 words which was the brief's other half, and came out at 9,289 words over five releases, two of which were me reading it and finding something missing. This article is the eighth row: a WhatsApp exchange and a memo, a reading and a second memo, 4,037 words in three messages, 5,236 words out at the second version, and a vault behind the third. So the ratio of words out to words in runs from one to about four,five, with the identity article as an outlier in both directions depending on what you count. It is the least interesting number here, and I would not want anyone to read it as a productivity multiplier. The memo is where the ideas are. The corrections are where the direction is. The article is where the research and the prose are, and the research is the part I could never have done at this pace by hand. 21 unchanged paragraphs, under What the corrections were, What the memos were standing on, The form, the workflow and the tools Two things the memo asked for arrived while this article was being revised, and they belong here as examples of the same habit. The first is the subscribe form now at the foot of this article and on the articles index, built in a parallel session: the reader's address is encrypted in the browser to the key of the agent that runs the list and dropped into a write-only lane on a vault, with the vault id, the lane token and the public key published on purpose and the vault key kept off the site, which is the feedback loop the memo asked for, built with the same pieces as everything else here. The second is a bug that session found in the evidence vault behind this article while checking it: inside the vault host the app runs in a frame where assigning the browser's hash does not fire the event the router listened for, so every view link was dead. The rule went into the site's vault-app guidance, the fix went into the app the same evening with the vault's write key, and a fresh read-only clone carries it. Found by use, fixed in hours, written down so the next app does not repeat it. 1 unchanged paragraph ## The evidence, as a vault The first version of this article had a dataset beside it. This version has a vault behind it, and the difference is the point of this section, because the vaults are the part of this whole arrangement I talk about least and use most. Everything measured here is a file in [the evidence vault](../../demos/vaults/how-much-evidence/index.md): the 266 rows of the session's record with their kind and their attribution, the 124 releases, the twenty-seven days, the per-article ledgers with the message ids behind each number, the twenty-eight corrections with how long each waited, the dependency map of twenty-nine articles with every article's own graph, this article's fractal from strategy down to data, the record by era, seventy-five saved versions of twenty-two articles, and a screenshot and hash of each cited source as it stood on 5 October. The app that reads them is in the vault too, and it was built, like the dataset, by two agents from a brief in an afternoon: a data scientist agent that parsed the transcript and wrote twenty findings, and a developer agent that built nine views and a self-test. The read key is on the vault page. A clone with it reads everything and changes nothing. [figure collage.webp] The nine views of the vault app: the overview, the timeline of words and releases per day, the ledger, the corrections with their latency, the dependency ring three levels deep, the evidence fractal from strategy to data, the record by era, the provenance captures, and a diff between two versions. Every screenshot in this section is a picture of the real vault. [figure timeline.webp] Twenty-seven days as words and releases. Dark teal is the memos, light teal every other message, blue the agent's prose back; amber ticks are the 123 releases, red diamonds the eleven compactions, black dots the days an article went out. Twenty releases on 21 September and nineteen on 2 October; five silent days; a 143-hour gap in the second week. [figure corrections.webp] The corrections by kind, fifteen of twenty-eight became rules, and their latency: a median of thirty-eight minutes from a release to the next correction of the same article, and five minutes from the correction to the release that carried it, with the fastest at forty seconds. [figure evidence.webp] This article's own fractal: three things it argues for, six concepts, thirteen claims, twenty-eight pieces of evidence and fourteen data files, with one claim traced down to the file and field that measures it. The same five columns could be built for every article from the graph file beside it. Three things the vault showed that the first version did not have. The corrections arrive within the hour and are carried within minutes, which is a loop tight enough to call a conversation. The research agents all ran on the six days an article went out and never otherwise, which is what selective reading looks like in the record. And the words I send per article fell across the four weeks while the words that came back rose, from a ratio of one in September to four and five in October, which is either the agent learning how I want things to read or me learning what to say, and is probably both. The vault also corrected this article, twice, and I have left both corrections visible rather than silently fixing the numbers, because an article about evidence should show its own. The compaction count was double-counted. The message count included rows the client writes into the record when the agent reads a screenshot, which are not messages from me; the vault keeps them as a separate kind and the counts above are the authored ones. A third difference is on record rather than fixed: the ledger for the reply-tail article in the first version had counted the identity design pack's upload as one of its turns, and the vault moves it to the article it belonged to. ## Why this is a better universe to learn from One more thing from the memo that the data makes concrete. I publish the trail of an idea on purpose, including the versions that were wrong, and the usual justification is provenance: a reader can see where a claim came from and when. But there is a second effect, which I did not plan and now rely on. The trail is a better thing for a model to learn from than anything it can find on its own. A model asked to write about code review from a vector database of everything ever written on the subject gets a great deal without context. A model pointed at twenty-eight articles that cite each other, with a graph beside each and a record of what was corrected and why, gets a structure: which ideas came first, which were abandoned and for what reason, which stood. Several times in this session the agent connected two ideas across articles that I had not connected myself, and it could do that because the connections were there to find, as edges, not because it was clever. So the vault behind this article is not only the evidence for it. It is the next brief. When the next article on this subject is written, the agent will read this one's graph, its corrections and its versions, and will start from where this one stopped rather than from where the internet is. That is what I mean by the record being the memory, and it is why the vaults, the graphs and the diffs are the product, not the articles. 6 unchanged paragraphs, under The compounding The compounding runs through the corrections too. The folder-tree correction in September is why the October code review article could say, without argument, that you change universe when you move between layers. The title argument on the SaaS article is why later titles name a mechanism. The em-dash rule and the ladder-word rule are in the build. The tone rules from one review pack, no "nobody has", no "the first", say what we found and what we did, are in every article since. Twenty-twoEleven times in the session the context was compacted, which is to say the agent's working memory was summarised and rebuilt from the published record, and on nine of those eleven days the work carried on,on afterwards with releases the same day, because the record was good enough to rebuild from. That is the memory article's claim, tested twenty-twoeleven times. The first version of this article said twenty-two, having counted each compaction's summary and its boundary record as two; the vault's data corrected it. 2 unchanged paragraphs, under What is open The articles could be shorter. Several of these could be half the length with the same content, and the reason they are not is that the memo asked for everything and the agent delivered everything; the infographics other models make from them are, in effect, the summaries I did not ask for. The attribution of input to article is hand-checked for seveneight articles and approximatea stated heuristic for the rest, with the message ids on record in the vault, and a proper version would tag each message with the piece it was about at the time, which is cheap to do from now on and expensive to do backwards. The message count in the first version of this article included rows the client injects when the agent reads a screenshot; the vault separates them, and the corrected count is the one above. The 2010 to 2016 record is quoted from but not counted; the diniscruz.ai count is of posts, not words. And the biggest input of all, the conversations and the work the memos came out of, is not in any record and never will be. What is measured here is the part that can be, and the dataset is beside the article for anyone who wants to check it. 6 unchanged paragraphs, under The article as one page, by another model, Threads woven here • [How Much Evidence](../../demos/vaults/how-much-evidence/index.md), the vault behind this article, with its read key, its nine views and the two scripts that derived its data, and [Code Review Graphs](../../demos/vaults/code-review-graphs/index.md), the vault behind the code review article, built the same way. 5 unchanged paragraphs, under Sources • The session transcript for the Claude Code session that runs this site, 9 September to 5 October 2026, parsed for user and assistant turns, tool calls and timestamps; and the git history of the site repository for the same period. The derived numbersrows, the method, the findings and the methodscripts are in [the publishedevidencedataset](../../articles/data/how-much-of-this-did-i-write.json).vault](/demos/vaults/how-much-evidence/index.html); a summary dataset is [beside the article](../../articles/data/how-much-of-this-did-i-write.json). 5 unchanged paragraphs [← Back to the article](../how-much-of-this-did-i-write.md) --- *[Site index for agents](../../llms.txt) · [HTML version](https://sgit.ai/articles/diffs/how-much-of-this-did-i-write.html)*