A booking failure that forced out a four-layer prompt architecture: turning what the brand must never say into a hard rule.
Stop Writing Tone Instructions. Layer Them. - Isadora Martin-Dye, Isadora & Co · Isadora Martin-Dye
21 min total·Actually worth watching closely: ~5 min·2 must-watch clips
- 0:00 – 2:40Listen
Stop writing the intern a longer briefing
Opens with the intern analogy to explain why making the system prompt longer and more detailed is a dead end. The real cause isn't that the examples are bad, it's that one prompt is forced to carry four jobs of completely different natures at once: to be situational, to be inviolable, to be expressive, and to check itself.
The single-prompt failure is coupled responsibilities, not wording, so more wordsmithing won't make it better.
This stretch is pure spoken argument and analogy. The speaker is standing there making a case with nothing on screen worth watching, so you lose nothing listening to it while doing something else.▶ Jump to 0:00Speaker · Isadora Martin-Dye - 2:40 – 7:01Skim
The four layers, and layer one
Pulls the stack apart: immutable identity, situational mode, example-anchored voice, post-generation veto. The weight lands on layer one, the things the brand can structurally never say, which have to live where the voice layer physically can't reach them.
Hard rules belong in a layer the voice layer can't touch, not in a more emphatic line of prompt.
A layer diagram comes up here; one glance at the layers and their order is enough, and the rest of the explanation follows fine by ear, with no need to freeze on any frame.▶ Jump to 2:40Speaker · Isadora Martin-Dye - 7:01 – 10:30Listen
A wedding venue and missing people
The same architecture reused in two opposite worlds: Bloom, the wedding venue assistant, and Threadline, a tool for families and missing people. In the latter, layer one simply forbids words like confirmed, matched and solved.
Saying "match" to the family of a missing child isn't a tone slip, it's the single most damaging thing this product could do, which is exactly why it has to be a hard rule.
The two cases are carried by the speaker's spoken contrast, with no interface or data to make out, so just listen for the tension in the cases themselves.▶ Jump to 7:01Speaker · Isadora Martin-Dye - 10:30 – 14:04Listen
The situational layer most teams skip
Layer two, the situational mode: the way maps routing folds live traffic into the calculation, work out who's in the car before the prompt runs. A venue coordinator can be told the limits straight, like a colleague; a couple planning a first wedding gets a different route. It also draws layer three's boundary, since soft context notes are only for setting tone and empathy and are never quoted verbatim.
The situational call has to happen before the prompt runs, instead of leaving the model to guess who's on the other side while it generates.
This stretch is conceptual distinction and paired examples, delivered at a steady pace with a coherent through-line, easier to follow on headphones than by staring at the screen.▶ Jump to 10:30Speaker · Isadora Martin-Dye - 14:04 – 16:40Listen
Confidently offering something that isn't real
The post-mortem on the ordinary failure that forced layer four into existence: the model was never given the calendar, but to sound like good service it warmly and specifically offered a date that was already booked. All three layers above were working exactly as intended.
Confidently offering something that isn't real is far worse than a cold refusal; the quality of the voice can itself become the fault.
The whole stretch is narrative. The speaker retells the exchange and the reaction, and the emotion and the detail are all in the words, so the visuals add nothing.▶ Jump to 14:04Speaker · Isadora Martin-Dye - 16:40 – 18:50Skim
From instruction to permission
How the layer-four veto works: read what actually came out, check it against real data item by item, and stop any date that isn't on the allowed list. The first three layers are probabilistic instruction; the fourth is deterministic permission.
The prompt is prevention and the veto is the check, and you need both, because the prompt will eventually lose.
The check flow is walked through against a diagram; seeing which step the output passes and what data blocks it is enough, since the spoken explanation already carries the full logic.▶ Jump to 16:40Speaker · Isadora Martin-Dye - 18:50 – 20:56Listen
Assembly order, and failing loud
Closing engineering practice: every surface assembles through a single entry point, hard rules first and tasks last, replacing the twenty-four freelancing prompts from before the refactor; the human context block deliberately renders ahead of the numeric constraints; and identity fields are never allowed a default. It ends on the product call to disclose the AI up front.
A missing identity has to fail loud. The old silent default had every venue emailing from another venue's address, and the trust eroded before anyone on the team knew there was a problem.
The closing practices are lessons learned plus a bet-shaped sign-off; the speaker says more off-slide than is on it, so listening gets you the most complete ending.▶ Jump to 18:50Speaker · Isadora Martin-Dye