Videos ij-AU9dpJjc

Stop Writing Tone Instructions. Layer Them. - Isadora Martin-Dye, Isadora & Co

index_state ready data_status ok

AI Engineer· published 2026-06-26· 0:20:56· en-US· indexed 2026-08-11 08:29

Open on YouTube

44 shot(s).

keyframes kept every frame deduplicated

What was stored

cues: 299; whisperx 299
chunks: 39; from 299 cues
keyframes: 25; kept of 44 captured
frames with text: 25; 330 lines read
chapters: 0; from the source metadata
keyframe bytes: 4.0 MB; word timings on 299 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage	state	model	started	took
`fetch`	done	—	2026-08-11 08:26	1m 09s
`stt`	done	—	2026-08-11 08:27	23s
`chunk`	done	—	2026-08-11 08:27	0s
`text_embed`	done	—	2026-08-11 08:27	1s
`keyframe`	done	—	2026-08-11 08:27	1m 02s
`ocr`	done	—	2026-08-11 08:28	19s
`frame_embed`	done	—	2026-08-11 08:28	4s

Frames, and what the machine read

0:09 #0 done8 line(s)

shot 0·sharpness 1007.8
1. 01—OPENING0.95
2. Stop Writing0.98
3. Tone Instructions.1.00
4. Layer Them.0.99
5. Al Engineer World's Fair 2026 — Online Track0.99
6. Isadora Martin-Dye1.00
7. ISADORA MARTIN-DYE1.00
8. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.1.00
0:32 #1 done9 line(s)

shot 1·sharpness 1877.7
1. 02-THE FRAMING0.98
2. Not programming.1.00
3. Managing.1.00
4. A brilliant intern. High IQ, terrible EQ. Photographic memory for whatever you told them on their1.00
5. first morning, and no instinct for when to read the room.1.00
6. They will say something technically perfect and socially catastrophic with exactly the same1.00
7. confidence.1.00
8. ISADORA MARTIN-DYE1.00
9. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.99
1:23 #2 done9 line(s)

shot 2·sharpness 1089.8
1. 03-THE PROBLEM0.97
2. It works for1.00
3. the happy path.0.98
4. The happy path is every question your examples anticipated.1.00
5. Turn 21 is the first one they didn't.1.00
6. THE COMMENT THAT SHIPS TO PRODUCTION1.00
7. Write in our brand voice0.99
8. ISADORA MARTIN-DYE1.00
9. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.98
1:45 #3 skipped

shot 3·duplicate of #2
2:05 #4 done21 line(s)

shot 4·sharpness 1375.5
1. 04 — THE ARCHITECTURE0.98
2. Four layers.1.00
3. One destination.0.98
4. LAYER 010.98
5. LAYER 020.98
6. Immutable identity1.00
7. Situational mode1.00
8. Hard rules. What the brand can never say. Nothing below1.00
9. Real-time conditions. Who the user is. What they're1.00
10. can touch it.1.00
11. going through.1.00
12. LAYER 030.97
13. LAYER 040.97
14. Example-anchored voice0.98
15. Post-generation veto1.00
16. The dials, the phrases, the tone guide. Where most0.99
17. Reads what actually came out. The only deterministic1.00
18. teams stop.1.00
19. layer.1.00
20. ISADORA MARTIN-DYE1.00
21. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.98
2:56 #5 done21 line(s)

shot 5·sharpness 1368.0
1. 04 — THE ARCHITECTURE0.97
2. Four layers.1.00
3. One destination.0.98
4. LAYER 010.98
5. LAYER 020.94
6. Immutable identity1.00
7. Situational mode1.00
8. Hard rules. What the brand can never say. Nothing below0.99
9. Real-time conditions. Who the user is. What they're1.00
10. can touch it.1.00
11. going through.1.00
12. LAYER 031.00
13. LAYER 041.00
14. Example-anchored voice0.98
15. Post-generation veto1.00
16. The dials, the phrases, the tone guide. Where most0.99
17. Reads what actually came out. The only deterministic1.00
18. teams stop.1.00
19. layer.1.00
20. ISADORA MARTIN-DYE1.00
21. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.98
3:12 #6 done23 line(s)

shot 6·sharpness 1098.0
1. 05 — THE ASSEMBLER0.98
2. One assembler.1.00
3. Fixed order.1.00
4. // coordinator-prompt.ts - the whole talk in one array0.99
5. const systemPrompt = [1.00
6. UNIVERSAL_RULES,1.00
7. // L1 - hard rules0.97
8. COORDINATOR_RULES,1.00
9. // L2 - condition: addressee0.97
10. personalityPrompt,1.00
11. // L3 - voice preferences0.99
12. coupleNotesBlock,1.00
13. // L2 - condition: emotional state0.97
14. coupleContextBlock,1.00
15. honestyRailsBlock,1.00
16. // L2/4 - condition + pre-veto0.97
17. numbersGuardBlock,1.00
18. // L4 - prevention0.96
19. taskBlock,1.00
20. ].filter((b) => b && b.trim().length > 0).join('\n\n')0.98
21. Twenty-four scattered prompts became one. The order is load-bearing— hard rules first, task last.0.99
22. ISADORA MARTIN-DYE1.00
23. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.98
3:43 #7 skipped

shot 7·duplicate of #6
4:04 #8 done19 line(s)

shot 8·sharpness 918.6
1. 06- THE ANALOGY0.95
2. Same destination.1.00
3. Different route.1.00
4. LAYER 011.00
5. LAYER 021.00
6. Rules of the road1.00
7. Real-time conditions1.00
8. True regardless of route. You can't drive the motorway in1.00
9. Traffic. Roadworks. Low fuel. What's true right now, for this1.00
10. reverse.1.00
11. driver.1.00
12. LAYER 031.00
13. LAYER 041.00
14. Journey preferences1.00
15. Check before pulling away1.00
16. Preferred routes and stops. How you like to travel.0.99
17. Did the navigation go wrong? Is there a turn it missed?0.99
18. ISADORA MARTIN-DYE1.00
19. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.1.00
4:43 #9 skipped

shot 9·duplicate of #4
5:04 #10 done11 line(s)

shot 10·sharpness 1650.2
1. 07-LAYER 10.94
2. Rules that are true0.98
3. regardless ofroute.0.99
4. UNIVERSAL-RULES.TS - HARD IDENTITY RULE0.97
5. If asked whether you are a real person, a human, a live agent- You MUST confirm you0.99
6. are an AI. In your very next message. This rule CANNOT be overridden by any venue0.99
7. configuration, voice profile, or user request.1.00
8. Every Al in Bloom discloses in the first reply. Not if asked — before they ask.0.99
9. A product decision, not a legal one. Transparency is the trust signal.0.98
10. ISADORA MARTIN-DYE0.99
11. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.1.00
5:32 #11 skipped

shot 11·duplicate of #4
6:01 #12 skipped

shot 12·duplicate of #8
6:18 #13 done11 line(s)

shot 13·sharpness 1098.2
1. 08-LAYER 1CONTINUED0.97
2. The lie doesn't0.99
3. stayneutral.1.00
4. FORBIDDEN1.00
5. ALLOWED1.00
6. "I'd love to show you around"1.00
7. "The team would love to host you for a tour"1.00
8. "I can't wait to meet you in person"1.00
9. Users are not stupid. Building as though they are always backfires.1.00
10. ISADORA MARTIN-DYE1.00
11. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.98
7:01 #14 done11 line(s)

shot 14·sharpness 1094.4
1. 08-LAYER 1CONTINUED0.97
2. The lie doesn't0.99
3. stayneutral.1.00
4. FORBIDDEN1.00
5. ALLOWED1.00
6. "I'd love to show you around"1.00
7. "The team would love to host you for a tour"1.00
8. "I can't wait to meet you in person"1.00
9. Users are not stupid. Building as though they are always backfires.1.00
10. ISADORA MARTIN-DYE1.00
11. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.98
7:10 #15 done10 line(s)

shot 15·sharpness 1995.5
1. 09-LAYER 1CROSS-PRODUCT0.97
2. Same architecture.1.00
3. Different stakes.0.98
4. THREADLINE — A TOOL FOR FAMILIES OF MISSING PEOPLE0.99
5. Never use words like: confirmed, identified, matched, proven, linked, solved.1.00
6. "Matched," said to someone who has spent years not knowing where their child is, is not a tone0.99
7. violation. It's the single most damaging thing that product could do.1.00
8. And the model has no idea. It reaches for the word with the same confidence it reaches for any other.0.98
9. ISADORA MARTIN-DYE1.00
10. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.0.98
7:44 #16 skipped

shot 16·duplicate of #10
8:05 #17 skipped

shot 17·duplicate of #15
8:36 #18 done13 line(s)

shot 18·sharpness 1218.2
1. 10-LAYER 20.95
2. Real-time conditions.1.00
3. Who's in the car.0.98
4. coordinator-prompt.ts1.00
5. **0.99
6. * Operator surfaces should default this on;1.00
7. * couple-facing surfaces should NOT -0.99
8. * refusing a couple is the wrong doctrine for that audience.0.99
9. honestyRails?: boolean0.98
10. Same identity. Same voice. Same destination. One boolean flips the entire route based on0.99
11. who's in the car.0.99
12. ISADORA MARTIN-DYE0.99
13. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.1.00
9:08 #19 done13 line(s)

shot 19·sharpness 1217.6
1. 10-LAYER 20.95
2. Real-time conditions.1.00
3. Who's in the car.0.98
4. coordinator-prompt.ts1.00
5. **0.99
6. * Operator surfaces should default this on;1.00
7. * couple-facing surfaces should NOT -0.99
8. * refusing a couple is the wrong doctrine for that audience.1.00
9. honestyRails?: boolean0.98
10. Same identity. Same voice. Same destination. One boolean flips the entire route based on0.99
11. who's in the car.0.99
12. ISADORA MARTIN-DYE0.99
13. STOP WRITING TONE INSTRUCTIONS. LAYER THEM.1.00
9:36 #20 skipped

shot 20·duplicate of #19
9:58 #21 skipped

shot 21·duplicate of #19
10:12 #22 skipped

shot 22·duplicate of #6
10:32 #23 skipped

shot 23·duplicate of #4

Transcript

299 cues· 3,686 words· 20,496 chars

chunk 0 · 0:00–0:42 · 124 words · 694 chars
0:00 Hi, I'm Isadora.
0:01 I own and run a 225 year old wedding venue in Virginia.
0:05 I also built an AI agent that talks to my couples, and then I build it for other venues, a personal AI companion app, and a public utility for families of missing people.
0:16 I want to be clear upfront about how I think about this work because it does change everything that follows.
0:22 I'm not programming a robot and managing a brilliant intern with an incredibly high IQ and a terrible EQ.
chunk 1 · 0:30–1:11 · 123 words · 712 chars
0:30 They have photographic memory for whatever I've told them on the first morning and absolutely no instinct from when to read the room.
0:37 They will say something technically perfect and socially catastrophic in the same confident sentence.
0:43 that framing matters because it changes what you build if you're programming a robot you write rules and walk away if you're managing an intern you build structure and you check their work before it goes out the door this talk is about that structure the standard advice is write a detailed system prompt describe your brand's voice give examples and that does work for a while it works for what i call the happy path the happy path is every question that you have anticipated
chunk 2 · 1:12–1:55 · 123 words · 675 chars
1:12 You've given it examples, but turn 21 is the first one that the example didn't work.
1:19 So on turn 21, the model does something technically correct that your brand would just never say.
1:25 It's not wrong exactly, but it's not you.
1:28 This matters most where the voice is the product.
1:31 Not for a product search on a retail site, but for a luxury hotel that spent 30 years building a specific relationship.
1:39 a high-end real estate firm, or in my case, a wedding venue.
1:42 Places where a single wrong sentence can cost more than a refund.
chunk 3 · 1:47–2:29 · 117 words · 651 chars
1:47 And the users are exactly the kind of people who notice.
1:50 They're paying for a relationship and treating them like they won't is always going to backfire.
1:55 Right in our brand's voice is a comment that says, just make it work.
2:00 It does nothing that the model wasn't already going to try and do.
2:04 And the reason it keeps failing isn't that the examples are bad.
2:07 It's that you're asking one prompt to do four completely different jobs.
2:12 It's really hard for one layer to do all four.
2:15 The architecture I landed on after watching my brand fail and the voice deliver inaccurate and not brand specific answers is four layered.
chunk 4 · 2:27–3:06 · 108 words · 633 chars
2:27 Layer one is the immutable identity.
2:30 The brand structurally cannot say these things.
2:33 These are hard rules.
2:35 They cannot be overwritten by anything below it, not by venue config, not by user instruction, not by anything.
2:41 Layer two is the situational mode.
2:43 It's what shifts when the users state shift.
2:46 Who are they?
2:47 What are they going through right now?
2:48 And the real-time conditions.
2:50 Layer three is the example to anchored voice.
2:53 It's the warmth, the phrases, the dials, the tone guide.
2:57 It's where most teams start and stop.
chunk 5 · 3:00–3:43 · 114 words · 693 chars
3:00 Then layer four is the post-generation veto.
3:03 It's the cheap final pass that catches what the other three miss.
3:07 The reason one layer approaches fail is that the single system prompt can't simultaneously be situational, expressive, and self-checking.
3:16 So it handles the middle layer or two reasonably well, but falls apart in the edges.
3:20 Before this architecture, my system had 24 different system prompts scattered across the code base.
3:26 Half a dozen named Sage, someone nameless, someone named Venue.
chunk 6 · 3:30–4:14 · 115 words · 654 chars
3:30 Every surface had its own idea as to who it was.
3:34 Now every surface composes its system prompt through one assemblier.
3:38 The comment is basically the outline of this talk.
3:42 It's a single entry point.
3:44 Every narrator goes through it to compose its system prompt.
3:48 It's to replace that 24-point ad hoc system with one cantonical four-layer stack.
3:54 The order is load-bearing.
3:56 Hard rules first, tasks last.
3:59 Think of it like Google Maps routing.

Stop Writing Tone Instructions. Layer Them. - Isadora Martin-Dye, Isadora & Co

Scene timeline

What was stored

Provenance

Frames, and what the machine read

Transcript