The talk, and the first deck
The talk was booked for a morning slot of twenty to thirty minutes with questions after, given over video to an audience the organisers described as junior staff to programme leads, mostly in small firms and local government, with a recording wanted as the backup. The person's first message, on Tuesday the 7th, asked the AI to gather everything it knew about his tool and then to draft a thirty-minute outline using a real project as the demonstration, aimed at an audience of leaders and not all "boring technical".
The first afternoon went the way such afternoons go. An outline, then mock-ups of the slides in a table, then his corrections one message at a time: the plausible requirements the AI had filled in silently on one project were never decided by a human; there was no slide on how records borrow from one another; the person who writes a requirement is not always the one who signs it; where were the five pains the talk should start from, the backlog nobody will delete, the things built that the client never asked for, every project started from nothing, the documents written at the end, the effort of checking they are right. A slide on what else a record can generate. A slide on his own earlier whitepaper's idea that the right behaviour should be the easiest path. "Intent persists. Tools and outputs change. Now turn this into an actual slide deck."
Five hours later he looked at the deck and said what he thought: "To me this is a pretty boring presentation. It's not clear how I present it. What the arc of the journey is, how I bring it to conclusion. The fact that there is a tool that does amazing things. The need to generate both a pre-recorded version which I would like to be more engaging and then the ability to present it live. It has no punch."
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. What was different here is what the record was of. A talk is not software. Its record is its argument: the one thing the audience should leave able to do, five acts each ending in what the audience should think, twenty claims the talk makes, the evidence each rests on, what the talk assumes about its slot and its audience, and six things it must never say or show. That record was written in one evening. The slides came a week later, on a second record the product could build from, and that one he chose not to sign at all; the first he signed in part on the same night the slides were drafted, and the signatures sit in a working copy nobody has committed.
A story, shaped and signed in an evening
The AI answered the "no punch" with a spine, five acts: open on a result that should not be possible, turn it into a worry the audience already lives with, give the one idea, prove it on a real record, then say what the record did not catch and land. "Yes, that's the right spine, write the full timed script." And then the sentence that made this an example: "I also think we should create an idd graph for the content, and two more graphs for the presentation, and then separate graphs for the demos. Let's make this an amazing demonstration in its own right."
The story's record was shaped first: why it needed an intent, acts, claims, evidence, assumptions, constraints and non-goals, and why it did not need requirements, tests, risks or non-functional rules, which a talk has no use for. Then the story itself: 55 statements. The five acts. Twenty claims, from "a one-page note became a live service with its legal paperwork in three days" to "a signed record does not show a decision was good; a person using the thing does". Thirteen pieces of evidence, each a dated fact from a record he holds. Six constraints, among them no resident's data on screen, the parking operator not named, a prototype called a prototype, every number quoted as the evidence records it, and no tool internals outside the demonstrations. Five things left out on purpose, among them any comparison with other requirements tools. None of it was signed that night. A week later, in the small hours of the 15th, the shape's 23 decisions and the story's intent, acts, constraints, assumptions and non-goals took his signature in his working copy, where they still sit uncommitted; the claims and the evidence he left proposed, and they still are.
Three of the claims are in his words rather than the AI's, and the record marks them so. Early the next morning he wrote that the method "is not useful everywhere": for a quick solution or a throwaway prototype the rigour never has a moment to return value, but on work many people will build on, that its author returns to, that is handed on, or that is reused across projects, "my experience is that the effort and rigour pays back". The AI wrote that down as a claim; he sent three corrections in ten minutes: "early" means before the first reason to have it, since you do not know when that will be; adding a record later is easy and sometimes easier, "it's human nature to put off things that are unfamiliar"; and the public catalogue of standards is left unsigned on purpose, so that an organisation that wants a signed version takes its own copy. Those three became the slide on when it is worth using and when it is not.
A demonstration built in an evening, and set aside
The same night he wanted a demonstration "in its own right", and chose one: a public tax calculator, held to the accessibility and security standards, driven by a rules engine, with the rules as a record of their own. By late that night there were eight records in a repository of its own, 511 tax rules among them, and he had signed the calculator's 76 decisions in his terminal and said he would sign every rule before the talk, "it is important as part of the demonstration". A challenge pass that night read the rules against the legislation and found the personal-allowance rule did not give the greatest reduction, which he sent back to be investigated. He also found that a signature did not yet cover an item's links, and had the tool itself changed first.
The tax rules were never signed and the calculator never reached the talk. When the slides were made a week later the demonstration had moved to a page that already existed, the calculator from the third example, signed in September. He put it on a branch, held it to the two standards it had never cited, signed the four new decisions, and then had the AI do on that branch the two things the talk needed to show: propose a change that crosses a signed line, and reword a signed requirement after its signature. Thirteen minutes, by the commit times. That branch is what the audience sees in the browser tools: a proposal that waits, a quiet change caught, every requirement traced to its reason.
The first draft of the slides recorded the switch, and he objected: "that looks plain wrong, why even mention my indecisiveness!" It went. A slide that used his own browser editor as a cautionary tale went too, "because it was my product that I failed to use throughline to develop", which is the fifth example on this site.
The talk as a deck
On Wednesday the 14th the talk became a series the product of the sixth example could build: a record of its own, in the same repository as the story, with the series, its audience, its topics, and for each slide its layout, its place and its words, and for each claim its source. The first deck built from it was the one the story had planned, opening on the question everyone adopting AI is already asked. That afternoon he pasted the title and pitch he had agreed with the organisers, which asked for something else: why work drifts from what was asked when AI is driven by prompts, and the tool as the answer. A second deck was written to that title beside the first, eight threads, and the first was kept.
What he asked for over the evening was the look and the motion. The three screenshots of the tools in the light theme, "the whole app around each screenshot"; the badge animated better than slide software could; the title on the first slide where the organisers' logo was, with his name and role, which the draft had left off; the calculator throughout, where the parking register had crept back. And a draft page: "can I see a draft, no motion, no subtitles, no narration, before I start signing, perhaps with item UID references." The product did not have one; it did by the next morning.
Thursday was the comments. In a fresh session he gave the AI a brief, the slide he would comment on, and then one message per slide: sixteen in the first 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 subtitle, a bullet, a title that had to fit one line. Introduce the tool before it is used; every decision captured, not signed, since two were not; "an AI makes a proposal that needs signing"; the sceptics' figures are about AI-assisted work, not the tool, and the slide must say so; when it is worth using and when it is not, with the bullet about wrangling documents added. Each comment became one commit on the record naming the slide. The deck rebuilt itself a few minutes behind him. Twice that day the product that built it got in his way, and both are in the sixth example: a build in the middle of a long narration pass, and a deck that dropped off the site when a slide broke a rule, "the slide has disappeared altogether :-(".
The figures on one slide went through three versions in twenty minutes. The draft said the calculator had zero links to the accessibility standard and fourteen criteria now met. "I'm concerned that 0 links to the accessibility standards and then 14 criteria. WCAG is accessibility standards!!!" Then: "How could it be 0. 0 is not impressive." Then: twelve criteria, not fourteen, and what a hundred per cent does and does not prove. The slide now leads with the forty decisions signed in September, none of them held to the standard, the missing link noticed in October, and closed the same evening. His own gloss on it, in the session: "it's also sad, because adding throughline sources gave us nothing new."
The voice, the words, and the night
The deck was spoken in his own cloned voice by the product's worker, and some words came out wrong: an acronym, and his surname. By that evening the narration text was editable per slide, and a tenant's rules for how words are said were on the product's record. Then he saw the subtitles carry the respelled words, and the product learned to keep what is written apart from what is said. At a quarter to ten he left the AI a list for the night with one rule, in his words: "take a note of the commit before you propose something that you signing and continue"; at ten to five he was signing from his phone.
Friday the marking: the briefing series' word stamped on a talk, and a cover line saying nobody had approved it. The second one is the decision this example turns on. The product's rule, signed on the 9th, is that a deck of a series that waits for sign-off is marked until every statement it shows is signed. He did not want the talk to wait: "Decks that don't wait: no marking anywhere, signed or not," by virtue of the deck's own approval setting. The product gained that setting that day. The deck's 167 statements are proposed, nobody has put a name to them, and the slides say nothing about it, all three by his decision.
The evening and the night after were content again. A slide that called the method's outputs "prototypes" when some are real documentation for a client; a slide titled "the honest part" that should instead say what to bear in mind: that a signed record is not automatically complete or correct, that the person who signs carries that responsibility, that a record should be challenged, that the tool has to be actively used. "Honestly, is it worth people using throughline or are the alternatives better?", at twenty past five on the Saturday morning, became a slide on where it sits among the alternatives, three spec-driven tools, behaviour-driven development and two requirements-management tools, with sources for each, and a line on what this one adds. And the number on the slide that says the talk was built the same way was corrected twice, once because the deck had 50 slides and not 35, and once because 89 signed items was the deck's count and not the product's, which was 967.
Saturday morning: drop an unfinished tool from the slide on the browser tools; say the browser tools are early access; "think of a better pair" for a sentence contrasting a council register with a software product, since a register is a product; a slide that named a ticketing tool out of nowhere; a screenshot retaken with the inspector panel open. At twenty to four the narrated video, the subtitles and the slide file were uploaded to send. The talk is on Wednesday.
Who did what
| Only the person could | Only the AI would | Both, through the record |
|---|---|---|
| Say the first deck had no punch, and ask for the story before the slides | Propose five acts and write 55 statements of story in an evening | Sign the shape, the acts, the constraints and the non-goals a week later; leave the claims proposed |
| Say when the method pays back and when it does not, in three corrections in ten minutes | Mark those three claims as his words and not its own | Turn them into the slide on when it is worth using |
| Choose a demonstration, build it in an evening, and set it aside | Read 511 tax rules against the legislation and find one that gave less than it should | Rework a signed page in forty minutes to show a proposal that waits and a change that is caught |
| Agree the title and pitch with the organisers, and hold the talk to them | Write a second deck to the agreed title beside the first, the same afternoon | Keep both decks on one record, their shared slides as twins |
| Comment on the slides one at a time, by number, from a brief he wrote | Make each comment one commit naming the slide, and rebuild the deck behind him | Correct the one slide's figures three times until they said what the record says |
| Decide that the deck does not wait for sign-off, and that the slides say nothing about it | Spend a night proposing and building without a signature | Add the decision the product's rule lacked, rather than weakening the rule |
| Ask whether the tool is worth using at all, at five in the morning | Answer with a slide of alternatives, each with a source | Leave the talk's own record unsigned, and say so here |
What went wrong, and what caught it
- The first deck was boring. Found by him, six hours in. Nothing could have caught it; a record of the argument was the answer, and it came next.
- The first demonstration never reached the talk. 511 rules proposed, none signed; a personal-allowance rule that gave less than the greatest reduction, found by the challenge pass; a signature that did not cover links, found by him and fixed in the tool. Set aside for a page that already existed and was already signed.
- The draft slides recorded the change of demonstration. "Why even mention my indecisiveness." Found by him on the draft page.
- The parking register crept back after the calculator had replaced it. Found by him; two commits to remove its last traces.
- A slide's figures were wrong three ways: zero links that were not impressive, fourteen criteria that were twelve, and a hundred per cent that proved less than it read. Found by him; the slide now leads with the forty decisions and the gap closed the same evening.
- A talk was marked with a briefing's word, and told its audience nobody had approved it. Both signed rules behaving as signed. Found by him on the slides. Fixed by adding the decision the record lacked.
- Empty placeholders on the cover that led with the deck's title. "Can you see the text boxes that do not need to be there?" Found by him from a screenshot.
- A rows slide with five items was refused by the builder, which allowed two to three, and the deck vanished from the site. The rule was right for a briefing; a talk needed five. Found by him; the rule now allows five, and the builder keeps the last good build, which is open.
- His surname and an acronym were mispronounced, then respelled in the subtitles as well as the speech, then one acronym still not respelled. Found by him, listening. Three changes to the product.
- A test voice with a nonsense name was in the deck's voice list and would have gone out. Found by him the morning after.
- A compare slide showed three items where the narration spoke four. Found by him, watching. The layout takes a fourth.
- Slides from forty-four on were silent in the presenter. Found by him; a refresh fixed it, and a deck rebuilt while on show is now taken up where it is.
- The "built the same way" slide miscounted twice. Found by him both times. Numbers on a slide about the record come from the record now.
- The subtitles in the video are not shown by a desktop media player. Found by him on the last morning. Open.
Time
| Stretch | What it held | The AI, active | The person, messages |
|---|---|---|---|
| 7 and 8 Oct | The outline, the mock-ups, the deck with no punch; the story shaped and written; the first demonstration built | 3.4 h | 57 |
| 14 to 17 Oct, the main session | The talk as a deck, the second deck, the demonstration on the calculator, the look and the motion, the night list, the content; shared with the product's own work | 26.2 h | 311 |
| 15 to 17 Oct, the comment sessions | Comments on the slides one by one, and the last morning's questions | 3.9 h | 80 |
The AI's hours are active time from the session records on the person's machine, measured by the assessment script with a ten-minute idle gap. The main session also built the product that made the deck, and its hours cannot be split; the sixth example counts the same hours. The story session's record was kept twice, and the two copies were merged before measuring. The messages are everything the person sent, including screenshots. The few times of day given are the UK clock.
Where it stops
The talk has not been given; it is on the 21st, and the recording is the backup. The story's twenty claims and thirteen pieces of evidence are still proposed, as they were on the 8th, and the 43 signatures on its structure and its shape have sat uncommitted since the 15th; the deck's 167 statements are proposed by decision; the two records do not cite each other, so nothing says which slide carries which claim of the story. The first demonstration's 511 rules are proposed in a repository nobody has returned to. None of the three records meets the assessment's "done", and no record is filed.