A field guide to Fable

Someone from the company that built the model, explaining how to use it. Which splits the talk’s credibility in two — the six techniques are trustworthy and the closing claim is the least trustworthy part of it.

The premise — “the bottleneck is me, not the model”

It opens on this line:

What limits you is not the model’s ability but your ability to find what you do not know you do not know.

To follow that you have to see that there are two kinds of not-knowing.

Kind	What
Knowing you don’t know	You know you don’t know this part — you can look it up
Not knowing you don’t know	You don’t know it exists — it never occurs to you to look

The second is the problem. Asking an AI requires knowing what to ask, and if you do not know it exists, you do not ask.

So the whole back half of the talk is methods for using the model to surface what you do not know you do not know.

Six techniques — three of which I was already doing

Technique	What you actually say
Blind-spot pass	I know nothing about this area. Find the things I wouldn’t know I don’t know
Several rough options	I have no eye for this. Give me four completely different directions on one page — I’ll react
Interview me	One question at a time. Ask the ones whose answers change the structure first
Point at a reference	This thing’s behaviour is exactly what I want. Read it and rebuild the same semantics
Log the deviations	When something pushes you off the plan, take the conservative option, write it down, and keep going
Quiz me	I want to know whether I understood this change. Build me a quiz I have to pass

The three that were new

One — “a third road that leaves a trace without stopping.”

When an AI hits an unexpected situation mid-task it normally does one of two things.

Usual	The problem
Stop and ask	You are away, and it sits still for hours
Decide and continue	Later, you cannot tell where it diverged

The third road: take the conservative option, write down where you deviated, and continue.

Not stopping means progress; writing it down means later you only have to read those lines.

Two — “I take a quiz before merging.”

The freshest idea here.

A check normally asks whether the code is right. This asks whether I understood it.

When an AI writes your code it is easy to pass it through without having understood it — it works. And later, when it needs changing, it is code you do not know.

So it has the AI examine you. Fail it and it is not yet time to merge.

Three — “sweep for blind spots before reading.”

Normally you look for what you missed after reading the code. This inverts it and sweeps before going in.

Being told this part is a complicated dead end and look at the history here first before you start changes the order in which you read.

But the closing section is the weakest part

At the end comes the claim “trade-offs are not real.”

Instead of weighing priorities and picking one thing over another, just do all of it and let reality show you the trade-off. Good, fast and cheap all at once, it says.

The grounds are one item: “I built this presentation deck in four hours the night before.”

That is weak grounds because a slide deck is cheap to get wrong. If it fails you redo it.

Where trade-offs actually bite is where reversal is expensive — a structure decided wrongly, a data migration done wrongly. Whether “just do all of it” holds there cannot be established by this evidence.

And this is a self-report about their own product, which is a different category from the six techniques above.

Which is why credibility splits inside one talk

Part	Why it is or is not trustworthy
The six techniques	The area the builder knows best. Nobody knows how to handle their model better than they do
The closing claim	A self-assessment of their own product, on a single piece of evidence

The same person, the same talk, expert in one part and interested party in another. They have to be read separately.

In fairness — the least saleable part was the most credible

Part three of the talk is not about gains but about loss.

Work that took weeks takes hours. How can you not smile, and how can you not nearly weep.

They loved the feel of writing code by hand and say they cannot go back.

Someone on the selling side talking about what their product took away is rare. And it is the least saleable passage in the talk.

Which makes it the most credible. It is not a line said to sell anything.

The detailed record starts here

Trust splits two ways inside the same talk. It is someone from the company that built the model explaining how to use it, so the six techniques are exactly the territory the makers know best and they hold up. The closing part’s “tradeoffs are not real” is a self-report about their own product, and its entire evidence is “I built this deck in four hours the night before.” Three of the six techniques turned out to be things I already do under other names, and three I didn’t.

What the talk says

The premise is one line — the bottleneck is not the model but my own ability to match map to territory and find my unknowns. So the whole back half is about using the model itself to surface them.

⚠ A limit first. I watched from 9:34 onward, out of 18 minutes. The first half is outside this document, and everything below rests only on the back half.

The six techniques:

Technique	Which box	What it actually asks for
Blindspot pass	Unknown unknowns	“I don’t know this module at all. Do a blindspot pass, find the relevant unknown unknowns, and help me prompt better” → it sweeps the module and flags gotchas like “this is a complicated dead end” or “go look at the history”
Brainstorms & prototypes	Unknown knowns	“I have no visual sense. Build one page holding four completely different design directions — so I can react to them”
Interviews	What I failed to state	“Interview me about the ambiguous parts, one question at a time — prioritise questions whose answer would change the architecture”
References	Known unknowns	“This behaviour in this library is exactly what I want. Read it and reimplement the same semantics on our side”
Implementation notes	What surfaced mid-run	“Keep a notes file. When an edge case pulls you off plan, take the conservative option, record it under Deviations, and keep going”
Quizzes	My own comprehension	“I want to check I understood this change. Build a report of context, intuition and what was done — plus a quiz at the bottom I have to pass”

The fourth is especially good. “The best way to give a map is to give another map” — the language and system don’t have to match. Want a React component? Hand over an HTML mockup as the map.

Three already, three new

The fifth is the big one. “When an edge case pulls you off plan, take the conservative option, record it under Deviations, and keep going” — one line that leaves a trace without stopping. Until now I had two modes: stop and ask (slow), or proceed and lose the reason later. There was a third.

The sixth I also didn’t do. The idea of being quizzed by the model on “did I understand this?” after the work is done hadn’t occurred to me. Code review looks at the code, not at my comprehension.

The first is familiar in shape but different in timing. Going into unfamiliar code, I asked questions after reading. I have never asked for the traps before going in.

What broke when I checked

The closing part is the weakest. “Tradeoffs are not real” — instead of prioritising this over that, do all of it and let reality show you the tradeoff. Good, fast and cheap: pick three.

There is exactly one piece of evidence — the slide deck built in four hours the night before. A deck is cheap to redo and nearly free to get wrong. The things I work on differ from that by orders of magnitude in failure cost. What holds for one class of artifact cannot generalise to “tradeoffs are gone.”

And this talk comes from the company that built the model. That is not a flaw so much as a reading instruction — the six techniques are the territory the makers know best and carry high trust; product-effect claims carry the least. Trust has to be divided inside a single source.

Verdict

What	Call
⑤ Deviations log	Adopt. The third path: leave a trace without stopping
⑥ Self-quiz before merge	Adopt. A slot that measures my understanding, not the code
① Blindspot pass	Adopt. Before going in, not after reading
②③④	Already doing. Two routes arrived at the same place
“Tradeoffs are not real”	On hold. Evidenced by one low-stakes artifact

One thing deserves fairness. Part three of the talk is not about gains but about loss. “You do in hours what might have taken weeks — how can you not laugh, how can you not almost cry?” He loved the feeling of turning a codebase over in his head and says he can’t go back. It is unusual for someone on the selling side to talk about what their product took away — and that passage, being the least saleable part of the talk, is the part I trust most.