The briefing, and the talk
The person is a consultant. Once a month he gives colleagues a five-to-ten-minute briefing on what has happened in AI that matters to their client work, in his employer's house style, and the September one existed as a slide file on one workstation. He also had a talk booked at an online conference for the 21st of October, with a title agreed with the organisers: how to get AI to do what you actually want, and a tool he had built for it.
His first message, on Thursday the 8th, asked for the October briefing in the house style, from screenshots of September's. Thirteen minutes later he changed the request: the content the AI had reached for was out of date, so instead of a deck he wanted a record of decisions, "so that we can be more deterministic and up to date", and the deck should come out of that.
That one sentence is what this example is about. Every other example on this site made one thing from one record. This one made a product whose job is to make things from records: a deck is a list of claims, each resting on a source, arranged into threads and slides, and a build turns the record into slides, a slide file, a narrated video and a page to present from. The product's own behaviour is on records too, eight of them. By the end, the talk was being built by a product that was itself being built, both from records, both by the same person and the same AI, in the same ten days.
How the work was divided
The arrangement from the other examples held. The AI proposes and nothing it writes counts until a named person signs it. Every decision says why it exists. Every change to the code names the decision it serves. The tool refuses to let a signature stand over words that have since changed. What was new here was scale and speed: eleven records across four repositories, 371 commits and 173 pull requests in nine days, the person often at a phone rather than a keyboard, and more than one AI session running at once, some of them in the cloud where he could not reach a terminal at all.
How the person signs was settled by him, in the record of the sessions, and had to be settled three times. From the first hour the hand-over gave him a command to run in his own terminal, an assistant that shows each item before asking for the signature, and he ran it at least ten times on the first day. On the fourth day he wrote it down as the contract: "I agree that I should do the ratifications. This is best done by you giving me a command of this form", that command, "wait for me to ratify. All the PRs and merges you should do once you have the ratified throughlines." Some sessions handed him a list to sign instead, which signs unread. On the seventh day he found he had "inadvertently ratified everything" when he meant to sign one item, and had the rest reverted. On the eighth day, at ten to five in the morning, handed a list for the second time that night: "I hate this ratification of multiple items without an opportunity to read", and within a minute the rule was the project's standing instruction to every future session, one record at a time, each item shown, never a list. Sessions in the cloud, where he had no terminal, got their signatures a third way, from a browser. The three ways, and how often each, are on the technical page.
One day: from an idea to a signed-in site
The record was shaped first. The AI proposed what the briefing needed to keep: who the audience is and what they care about, the topics, the sources with when each was checked, the claims each source supports, the threads that string claims into a story, the slides, the editions, and the rules every edition must pass. The person agreed the shape in two messages and signed its 36 decisions twenty-nine minutes after his first message. Seven minutes later he signed the rules themselves, 75 statements: a claim rests on a source; a news claim rests on a source published inside the edition's window; every thread carries news; a slide that states a number shows its claim; the same record gives the same deck; one command checks an edition and names what failed.
Then the October edition. A research pass found candidate stories; the AI read each page and recorded the sources and the claims they support, and he corrected what he saw on the slides as they were built: make the vendor explicit on both cards, add the one story to the right thread, add speaker notes. The edition was signed and built. He asked for it to be published behind the estate's sign-in, with a list of editions, a presenter in the browser and a download, and twenty-six minutes later it was. The site's own record, eighteen statements, was signed an hour and a quarter after the site was live: the one place in the ten days where the code came before its record. Seventeen signing sessions that day, one every half hour or so, each in his own terminal with the assistant that shows an item before it is signed, and most followed by the same four words: "all ratified, commit and push".
In the evening he asked for the thing that turned a script into a product: a way to run more than one series, and a way for an administrator to create one. The AI asked what a series is, where it is created and what "manage" covers, as a short list of options; he answered in letters. A record for the service was shaped that night, with the shape record widened to say why, and it was held to a published security standard at its second level, 263 of whose clauses it now cites. He signed 107 decisions about the shape of the records alone over the ten days.
Five training decks in a day, and a presenter in two windows
Friday began, before nine, with "is there any way to have the presentation run in one window whilst the notes and the navigation can be in a second window", which was on the record, signed and built within forty minutes. Later that day he asked for a second kind of series: training decks, one per consulting role, each answering what AI can do for me, why I should use it, what my responsibilities are and how it works with the tools of my trade. A roles register was added to the record. Then, one message each, "do the business analyst deck next", "do the developer deck next", the tester, the architect: five decks written from sources and signed one after another through the rest of the day and into the night, the last at half past ten, each deck's claims read and signed before it was published.
That evening he asked whether the decks could be video, narrated from the speaker notes by a speech service, stopping after each chapter, and holding a slide long enough to read it when there are no notes. The AI set out the choices as A, B and C. "A, B and C, YouTube later." The rules for narration, the pause at the end of a thread and the video downloads were on the record and signed by nine that night.
The build that spoke every slide
The first narration was done the wrong way round, and the record says so. The build that made the site also spoke every slide of every deck through the speech service before anything was published. A build took twenty minutes, and the last full one an hour and seven, and it ran again for every deck whenever anything changed. The speech service, shared with other things, had been killed for memory under it. The person's handover for the next session put the phases in order: the build publishes slides and a slide file and never speaks; a worker, minutes later, makes a silent video that is permanent; then each note is spoken once and kept under a name made from its words, the voice, the model and a seed, so that a changed note re-speaks one slide and not a deck; the page shows how far each deck's narration has got; and if the speech service is down the number stalls and the page says so. Six rules were changed and eight added, with a test for each, signed, and built that day.
Two more decisions the same weekend were about names, and both were his. The product had been named after its first series, and the series after the briefing; "one name per concept", he said, so the product was renamed and the series kept its name, through the repository, the images, the namespace, the host and every variable, with the old address redirecting. And the voice the narration should use was his own, cloned in the speech service, which belonged to him and not to the organisation the product was speaking for; so the speech service gained the idea of a voice granted by its owner to a named tenant, on a record of its own, before the product could use it.
Then the sharpest message of the ten days, on the Saturday evening, when he found the product was still a single static page. He had said, more than once, that every service in the estate should look and behave like its siblings: one consistent interface, a super-administrator, tenants with their own administrators, the management of series inside it. "This needs to be clear from the throughline graphs. This shouldn't be up for debate." It had not been on any record, so it had not been built. The tenants, the roles and the estate's shared sign-in were put on the service's record that night and signed; an application with pages was shaped on a record of its own, held to the same security standard and to the accessibility standard, begun that night and built on the Sunday.
His own voice, and rules for what the machine says
Sunday morning, in his words, the product was "not very responsive, and very difficult to use because of delays". The decks' files were on a slow volume; they were moved, and the browser was told it could keep them. The presenter had gained a flicker between slides; a faithful port of the old one removed it.
Then the voice-over. He wanted to record his own narration slide by slide in the browser, several attempts at each, choose the take, trim it at the start, end or middle, and have the videos made from his choices, with a button in a live presentation that plays his recording for the slide on show. The AI put five design questions to him first; he answered with letters. By the afternoon the recording and the takes were on the record, signed and built. That evening he asked for a proper take editor: hear the sound as the cursor moves, click to set the start and end, drag the edit points, zoom, see the frequencies as well as the amplitude, erase an artefact in the frequency view, and record a stretch again and splice it in. "Let's generate the throughline graph entries first." Then, having read them: "should these requirements not link to a specific component page feature on the application?" They should, and they do; a requirement for a page now names the page.
And the same afternoon, having listened to the first narrated deck, he noticed that a speech service reading speaker notes aloud misses what the notes mean. He pasted one slide's notes as an example. Should the narration be written for the ear, from the notes and the slide, rather than read from the notes? "It can when framed and available in the slide. We should build it so that it can see the slide." A fourth record was shaped for the rules of narration: sixteen rules, eight safeguards, three worked examples, a test for each rule, and the worker was changed to write each slide's spoken script by those rules before speaking it. He signed them within the hour.
Subtitles followed on the Monday, timed by a transcription service's word times, and his own takes transcribed so that their words become their subtitles. Several of these pieces were built by sessions in the cloud, where the person could sign only through a browser. On the Tuesday he noticed that three signing sessions on three records in the same repository had started three builds of the product, "none of which is needed since these are graph-only changes", and two rules went into the pipeline that morning: a change to a record alone starts no build, and every change to a record runs the tool's check before it can merge.
Decks shared between organisations, and series in their own repositories
On the Tuesday evening an error on the editing page turned into the week's largest decision. The AI offered a quick fix and a long one. "I want the long term solution. Decks can be associated with multiple tenants and the styling changes by tenant." He was away from his computer, controlling the session remotely, and could not sign: "can you go through these with me and then sign on my behalf with my authority?" The AI read him each item and he corrected two of them: re-branding is mandatory, not optional, and both the super-administrator and the owning tenant's administrators may share. No signature was given by proxy. The requirements were pushed with a note that ratification was asked for again, and he signed them himself the next day, in his terminal. On the first day, asked to "ratify the builder graph too", the AI had answered that a signature has to be his; it held to that.
The next morning he looked at how long a build would take now that every deck was made in three brands, and asked why the decks' records lived inside the product at all. "I think we should support having external throughlines to define decks." That day a series became a record in a repository, branch and path of its own, configured in the application by a tenant's administrator; the two existing series were moved out, 599 statements of the product's own record retired in one commit and the rules left behind for every series to borrow; and a builder in the cluster watches each series' repository and rebuilds its decks when it moves. A series proposed inside the application is written into its own repository as a pull request, which is how the product's own front door leads back to the record.
Then: "when presentations are being successfully built from external throughline graphs it is important that only the slides affected are rebuilt", with the narration, the recordings and the video either marked stale or left alone. The AI asked two questions about what counts as changed; "1a, 2a". Only what changed is made again, from that day.
The same day brought the complaint that shaped the rest of the week: "I am running out of time the iteration loop does not match my way of coding. I cannot wait 40 minutes each time we fix one thing!" The product's three images were being tested one after another on a busy runner. By the end of the day the tests ran at once, on a priority runner; by the Thursday evening, "under pressure and need quick turnaround", the builds moved to a hosted runner service, and a build of the main branch fell from about eighteen minutes to about five.
The talk
On Wednesday the 14th the talk began, on a record of its own in a repository of its own, a series like any other to the product. It had an agreed title and pitch, an audience the organisers had reported, a demonstration to show, and screenshots of the tools to take. The person's instructions were specific and came fast: use the light theme for the tools' screenshots, show the whole application around each one, animate the badge better than slide software could, drop a slide that used his own product as a bad example, "and I think we can do much better with the motion when moving from the first slide." Then he asked for something the product did not have: "can I see a draft, no motion, no subtitles, no narration, before I start signing, perhaps with item UID references." A draft page was on the record and built by four the next morning: every slide as a still beside the statement it comes from, whether that statement is signed, its notes, the claims it rests on and their sources, and an appendix of what no slide shows. He signed it at four, again at a quarter past, and again at five, because other work merging ahead of it renumbered its statements twice.
Thursday was the longest day: sixteen and a half hours of the AI's active time and 233 messages. The worker was making every deck in turn and his talk was behind six training decks in the queue. "Didn't I say the focus is only the stop writing prompts?" The queue became a page: both queues shown, decks put first, held, released, and "only this one". Four more fixes to the order followed in the day, each a rule on the record, before the product reliably finished one deck before beginning the next. A build that landed in the middle of a long narration pass renumbered the slides under it; a deck refused at a new revision dropped off the site instead of keeping its last good build, "the slide has disappeared from the presentation manager altogether :-("; the silent video of a deck with motion took over an hour because every frame was written as a picture first. Each is an issue, each with the reason and the fix proposed, two of them still open.
In a second session that day he commented on the slides one at a time, by number, and the AI changed the talk's record: sixteen comments on the slides in ninety minutes, from "we do not need to mark it as OFFICIAL in the footer and header on every slide" to the exact wording of a title that had to fit on one line. Each became a commit naming the slide. That evening he needed to edit what the narration says, because an acronym and his own surname came out wrong, and could not: the spoken script was written fresh by a model on each pass and a good wording could not be kept. By seven the narration text was editable per slide and audible in seconds; by eight a tenant's rules for how words are said, an acronym spelled out, a name respelled, were on the record. Then he saw that the subtitles now showed the respelled words. The script was split in two, one written for the reader and one said for the listener, and the subtitles follow the first.
At a quarter to ten that night he left a list for the AI to do overnight, with one rule, in his words: "take a note of the commit before you propose something that you signing and continue. Then if I ratify everything then we are good and if I don't we know where to start from". Six changes went through that night, each proposed on the record, built and tested; five waited on a signature. At ten to five in the morning, from his phone: "if you give me links I'll take a look at any items that need ratifying." He signed them all by twenty past five and the five merged.
Friday the 16th was the marking. "It still says OFFICIAL on the slides!!!!": the marking a briefing series wants was being stamped on a conference talk, so the marking became the series' own word, and a talk has none. "Argh, why does the first slide say 'Not yet approved by a person'!!! I don't want that on the presentation." That one is the example's best lesson, and it is below. Later that day the narrated video was checked clean: no marking, no approval line, no underlined links, the right words in the subtitles, his wordings. He spent the evening on content, the wording of a slide about what a signed record does and does not prove, an introduction for the demonstration he had been referring to without introducing, and at twenty past five on Saturday morning, "Honestly, is it worth people using throughline or are the alternatives better?", a question for the talk and for this page.
Saturday: presenting on one screen while controlling from another, so a live demonstration can hide the slides; and at half past ten, "when a new slide deck is generated and completed is it possible for us to auto refresh any presentations that are live?" On the record, built, tested and merged in twenty-seven minutes. At twenty to four the files were uploaded to send. The talk is on the 21st.
Who did what
| Only the person could | Only the AI would | Both, through the record |
|---|---|---|
| Ask for a record instead of a deck, thirteen minutes in | Shape eleven records, 1,821 statements, every one written by it | Sign the rules before the first edition was written to them |
| Decide that the product must look and behave like its siblings, and say it was never up for debate | Turn one slide's notes into a case for narration rules, and write sixteen rules with a test each | Put the tenants and roles on the record the night they were asked for, before any page existed |
| Say "one name per concept" and rename a running product | Read 81 sources against 228 claims and find fourteen claims saying more than their source | Move two series out of the product into repositories of their own, retiring 599 statements in one commit |
| Write the signing contract, then sharpen it when a session broke it | Work through a night on a list, proposing on the record and building, without a signature until morning | Sign thirty-four moved statements again, each reported by the tool |
| Comment on fifty slides by number, and decide what the talk says | Measure a build, a narration pass and a frame encode before changing the order of the work | Keep the talk's own record unsigned, and make the product say nothing about it |
| Sign from a phone at five in the morning | Propose a draft page for reading a deck beside its record, the day it was asked for | Record each fault as an issue with its cause, including the three still open |
What went wrong, and what caught it
- The build spoke every slide of every deck. An hour and seven minutes, and a shared speech service killed for memory. Found by the person, from the build times and the service's death. Fixed by putting the phases in order on the record.
- The product was a static page when the estate's rule said otherwise. The rule was in the person's head and in other products, not on this record. Found by him, on the third day, loudly. Nothing on the record could have caught it, because it was not on the record; it is now.
- A signed rule stamped a conference talk with the marking a briefing wants, and a line saying nobody had approved it. Both behaved exactly as signed: a deck that waits for sign-off is marked until it is signed. The talk did not wait for sign-off, by his choice, and the record had no way to say so. Found by him on the slides. Fixed on the record: the marking is the series' own word, a deck may set its own approval, and a deck that does not wait is marked nowhere.
- Three signing sessions started three builds. Found by him. Two rules in the pipeline the same morning.
- Fourteen claims said more than their source. Of 228, read against 81 sources by the AI in a pass he asked for. "Confirmed the fixes" where the source said a third party had tested them; "must be protected" where the principle said "consider how you protect". Amended on the record with what the source says, and the six slides that repeated the wording.
- The queue starved the one deck that mattered. Found by him, five times in one day, each a different rule about order. Caught by nothing; the order of a worker's pass was never on the record until it was.
- A long narration pass read files a newer build had renumbered, and a deck refused at a newer revision dropped off the site. Both found by him from the symptoms, both written up by the AI as issues with the cause, both open.
- Every frame of a moving slide was written as a picture before the video was encoded, and the silent video took over an hour. Found by the AI with a profiler when he asked why nothing had happened in seventy-five minutes.
- The subtitles showed the words respelled for the speech service. Found by him. The script was split into what is written and what is said.
- A test voice named after a nonsense word was in the deck's voice list, and would have gone out in the video. Found by him, the morning after it was added.
- "92 slides?!!!!! There are only 50." Progress counted the thread videos as slides. Found by him; fixed and signed within the hour.
- One signature given in bulk when one item was meant, and a list handed over twice in one night. A list signs unread. Found by him at once each time; the bulk signature reverted and the one item signed alone, and the rule to hand over the command that shows each item is now the project's standing instruction.
- The subtitles in the video are not shown by a desktop media player. Found by him on the last day, and open.
Time
| Day | What it held | The AI, active | The person, messages |
|---|---|---|---|
| 8 Oct | The shape, the rules, the October edition, the site, the service's record | 6.3 h | 53 |
| 9 Oct | Two windows; five training decks; narration and video on the record | 6.8 h | 42 |
| 10 Oct | The worker; a tenant's voice; the rename; tenants and roles; the application begun | 7.6 h | 88 |
| 11 Oct | Speed; the presenter; voice-over and takes; narration rules; the take editor | 7.6 h | 136 |
| 12 Oct | Subtitles; the take editor finished; mostly in the cloud | 0.7 h | 9 |
| 13 Oct | Graph-only changes; the check in the pipeline; the long-term decision, signed by proxy | 2.6 h | 19 |
| 14 Oct | Brands and sharing; series in their own repositories; incremental builds; faster tests; the talk begun | 11.5 h | 120 |
| 15 Oct | The queue; the draft; slide comments; narration text; how words are said; written and spoken; the overnight list | 16.7 h | 233 |
| 16 Oct | The marking; the approval line; progress; the video checked clean; the talk's content | 6.5 h | 89 |
| 17 Oct | Remote presenting; live refresh; the upload | 4.9 h | 73 |
The AI's hours are active time from the fourteen session records kept on the person's machine, measured by the assessment script day by day with a ten-minute idle gap, and come to 69.7 hours over the whole window. They leave out the sessions run in the cloud, which merged eighteen pull requests and left no record here, and the sessions that built the speech service's tenancy, the shared sign-in and the graph explorer the talk shows, which were adjacent work. The messages are everything the person sent, including pasted screenshots and the automated notes a watched pull request sends. The few times of day the text gives are the UK clock, and all of them fall at the ends of the day.
Where it stops
The work is not at the point the assessment script calls done, and no assessment record is filed for it. Two pull requests are open. Eleven requirements across three records have no test yet; one signed test's wording has moved and waits to be signed again; two requirements were deferred without ever being signed. The twelve training decks hold 220 statements nobody has signed, seven of the decks having been written in a cloud session and never read. And the talk itself, 167 statements, is signed in exactly one place, the line that says the series is live: the person chose that the talk should not wait for sign-off, "by virtue of the Deck's approval setting", and the product was changed that day so that a deck can say so. The record of what the talk claims and where each claim comes from exists; nobody has put their name to it.