Low-poly pile of crumpled drafts and broken prototypes feeding a glowing path into one clean desk display

Quick answer

Give AI your failed attempts means handing a capable model every earlier version you rejected, plus your “no, that’s wrong” feedback, and asking it to work out what you actually want before it builds anything. In a new video, Nate B Jones says Claude Fable 5.2 did exactly that for an e-ink desk display he had failed to get right for over a month, across three earlier attempts and several AI tools. It read hundreds of files, proposed a design he liked straight away, then built it. The method works with any strong model, but Fable 5.2 is expensive, so save it for problems where a good answer is worth real money.

Key takeaways

  • Your failures are context. Rejected drafts and your complaints tell a model a lot about what “right” looks like.
  • Ask it to infer, then propose. Get a design or plan you can approve before any building starts.
  • Split design from build. Fable 5.2 made its own handoff file and suggested a fresh chat when the first one got full.
  • You are the judge. Nate is not a designer, but he could recognise the right answer because he had thought hard about it.
  • Pick the job carefully. API pricing is $8 per million input tokens and $40 per million output tokens, so use it where the payoff is clear.

🛠️ Practical guides

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

Most of us treat a failed AI attempt as wasted time. We close the chat, start again and hope the next prompt lands. Nate B Jones suggests a different habit, and it is one of the most useful ideas we have seen from the early Claude Fable 5.2 coverage.

We have already covered what Fable 5.2 is and how it performs, and which Claude model suits which job, so we will not repeat that here. This article focuses on what is distinctive about Nate’s video: how he used a month of failed attempts as the brief, and how to decide whether a job deserves an expensive model at all.

Watch: the source video

What this video covers: Nate B Jones explains how Claude Fable 5.2 read a month of his failed e-ink display attempts, inferred what he wanted, wrote its own handoff file between design and build, and why the model’s price means you should pick its work carefully.

Video: Nate B Jones (@NateBJones), “Claude Fable 5.2 read a month of my failed attempts and got it right the first time.”, published 12 November 2026. No paid sponsor is mentioned; Nate points viewers to his Substack newsletter for the full write-up. This summary, the opinions and the practical advice are our own.

What happened when Nate gave Fable 5.2 his failed attempts?

Nate wanted an information display for his desk, built on an e-ink screen. He describes himself as picky about design: which information appears first, what stands out and how it all fits together. He had tried for more than a month, with three earlier attempts across different AI labs, models and tools, including the previous Fable 5.1. None of them satisfied him.

With early access to Fable 5.2, he tried something different. He gave it all of his earlier attempts. He admits his instructions were mostly blunt: it doesn’t work, fix it. But he also told the model, openly, that part of the job was to read between the lines of the failed versions and his earlier feedback, where he had said “no” a lot, and to work out what he had been pointing at.

The result, in his words, exceeded his expectations. Fable went through hundreds of files, came back with a design he liked immediately, and then built it. The finished display changes through the day to show what matters in the morning, afternoon and evening, previews the next day, shows the weather and rotates quotes from authors he reads.

💡 Tip: Nate’s own point is that the e-ink screen is not the story. The story is the kind of job: something you have thought about a lot, can recognise when it is right, but struggle to describe up front.

Why does giving AI your failed attempts work?

A lot of real work is “I’ll know it when I see it”. Editing, design, page layouts, email sequences, product names. You can spot what is wrong far more easily than you can write a perfect specification.

Each failed attempt quietly records that knowledge. Every rejected draft says “not this”. Every complaint, however vague, points towards something. On its own, one “no” tells a model very little. A month of them, read together, starts to draw the outline of what you want.

What you hand overWhat the model can learn from it
Earlier drafts or builds you rejectedWhat “wrong” looks like in practice, not just in theory
Your feedback, even blunt or vaguePatterns in what you keep pushing against
Anything you kept or half-likedWhich parts are worth saving
Your goal in one or two sentencesThe point all of it is meant to serve

The hard part, as Nate says, is getting from “I know this isn’t right” to “here’s what we should build”. That is the step a strong model can now help with.

How do you give AI your failed attempts? A 6-step method

You do not need early access or a big budget to try this. Here is a simple version for any project with a trail of near-misses.

  1. Gather the trail. Put every earlier attempt in one folder: drafts, screenshots, code, old chat exports. Do not tidy them up; the mess is useful.
  2. Collect your feedback. Copy the comments where you said what was wrong, even if they are short. If they are scattered, paste them into one document.
  3. Say what the job really is. Tell the model plainly that you want it to infer your intent from the failures, not just try again.
  4. Ask for a proposal first. One design or plan, with a short list of “what you want” and “what you don’t want”. No building yet.
  5. Judge it honestly. If it is close, say exactly what to keep. If it is wrong, use our guide to giving AI feedback when you can’t say why.
  6. Build in a fresh chat. Once the design is approved, move to a clean conversation with a handoff file (more on that below).

Here is a starter prompt you can adapt:

I have tried to build this [project] several times and none of the attempts were right.
Attached are all my earlier attempts and my feedback on them.

Part of your job is to read between the lines:
- Look at what I rejected and what I said was wrong.
- Work out what I was actually pointing at, even where I said it badly.
- Tell me, in a short list, what you think I want and what I clearly do not want.

Do not build anything yet. First propose one design and explain how it fixes
the problems in the earlier attempts. Ask me up to three questions if you need to.

⚠️ Warning: Check what is in the folder before you upload it. Old files often contain client names, passwords, API keys or personal data. Remove anything you would not paste into a public chat, and follow your own data rules.

Why did Fable 5.2 ask for a fresh chat?

As the design conversation grew, Fable 5.2 noticed its context window was getting heavy, nearly full. Without being asked, it wrote a handoff file and suggested starting a new chat for the build. In effect it said: design is done, I will do better in a fresh thread, here is everything you need.

Nate picked up the file, started a new chat, asked it to build, and the build worked. He describes the flow as Design → Handoff → Build. He also notes he has recommended this as good practice for a long time; it is just nice to see the model suggest it.

You can do this yourself with any AI tool. Our step-by-step guide shows how to write an AI handoff file and move to a fresh chat, with a template.

What is your job while the AI does the work?

This is the part we would underline. Nate is clear that this was not “cognitive surrender”, giving up and accepting whatever came back on the fourth try. He had spent hours thinking about the display, so he could tell when it was finally right.

He says he is not a designer. He could have proposed designs, but he was better as a judge, acting like the customer. His wider advice is to know where your own strengths are and engage with the model from there. He frames it as moving from AI as an assistant to AI as a cognitive partner: you still think hard, still decide and still communicate what you are aiming for.

  • Recognise useful problems worth handing over.
  • Judge proposed solutions against what you actually need.
  • Explain what needs to change, even roughly.

Those three skills, Nate argues, will decide who gets real value from these models, even when everyone has access to the same one.

What does the evidence say about Fable 5.2?

Separate the evidence from the claims here.

ClaimSourceHow to read it
Fable 5.2 completed 41.3% of Remote Labor Index projects at least as well as a human professional, vs 21.3% for Opus 5.5Anthropic’s own system card, as shown in the videoCompany-reported benchmark of 240 real freelance projects; not independently checked by us
It inferred his design intent from failed attemptsNate’s own experienceOne person, one project, with early access
Someone spent $250 on one API conversation about strategy documentsNate’s first-hand accountAnecdote, but a useful warning
A test run paused while OpenRouter waited for more creditsReported to NateAnecdote
Superhuman AI researchers may arrive next yearThe authors of the AI 2027 projection, as Nate reportsA forecast and opinion, not a fact

Nate himself says the benchmark number matters less than the size of the jump. Treat it as a signal that the model is worth testing on hard work, not a promise about your results.

Is Fable 5.2 worth the cost for this kind of job?

Nate offers a simple way to think about it. Every task falls into one of three kinds:

Kind of taskWhat it meansWhat to do
1. Already handledToday’s model does it well. Nate calls this task-level saturation.Stay on your cheaper model. An upgrade adds little.
2. Better, but hard to justifyThe new model is noticeably better, but not by enough to pay for.Use it sparingly, or not at all.
3. Hard enough to be worth itA job that is difficult and valuable to get right.This is where Fable 5.2 belongs.

On price, Nate quotes the standard API rate of $8 per million input tokens and $40 per million output tokens, lower than Fable 5.1’s $10 and $50 but still a lot for everyday use. He agrees with Anthropic’s advice to start with Opus 5.5 for most coding and knowledge work and only switch up when the task justifies it.

If you are on a subscription plan, using Fable within your weekly allowance costs nothing extra. The video mentions a launch promotion allowing up to 50% of weekly usage limits on Pro and Team standard seats; check Anthropic’s current terms, as promotions change. If you pay per token, Nate’s test is simple: would spending $300 to $500 solve a problem worth around $10,000? If yes, it is an easy call. If not, use a cheaper model.

💡 Tip: For a fair way to compare models on what a finished job really costs, see how to compare AI models on token cost.

What should small teams do about access?

Nate says he knows companies taking both approaches: trusting staff with access to the expensive model, or locking it off. If you give people access, give them two things:

  • Clear guidance on which projects are worth frontier-model spend.
  • Clear routing that sends everything else to a cheaper model by default.

For a one-person business, that can be a single sticky note: “Fable only for type-3 jobs. Everything else on the everyday model.”

🔎 Unconfirmed / rumour: Nate closes by arguing that AI labs point their best models at building the next generation of AI, so what we see in public is a small fraction of what exists inside the labs. That is his view and a common industry argument, not something the video proves. His practical advice is to focus on the work in front of you.

Frequently asked questions

What does “give AI your failed attempts” mean?

It means sharing your earlier rejected drafts, builds and feedback with an AI model and asking it to work out what you actually want from them, instead of starting again from a blank prompt.

Do I need Claude Fable 5.2 to do this?

No. The method works with any capable model that can read files. Nate found Fable 5.2 was the first to get his particular project right, but a cheaper model is a sensible place to start.

How much does Claude Fable 5.2 cost?

Nate quotes $8 per million input tokens and $40 per million output tokens on the API. On a subscription plan, use within your weekly limit costs nothing extra.

What is task-level saturation?

Nate’s term for jobs your current model already handles well, so a more powerful model adds little value for the extra cost.

What is an AI handoff file?

A short document that sums up a long AI conversation, including decisions, files and next steps, so you can continue in a fresh chat without losing context.

Is it safe to upload old project files to an AI?

Only after checking them. Remove passwords, API keys, client details and personal data first, and follow your own data and privacy rules.

The bottom line

The most useful lesson from Nate’s video is not about one model. It is that the drafts you threw away still hold valuable information. Gather them, be honest about what is wrong, and ask a capable model to infer what you want before it builds. Then stay in the judge’s chair. Save the most expensive model for jobs where getting it right is worth real money.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.
One thought on “Give AI Your Failed Attempts: How Fable 5.2 Got It Right First Time”

Leave a Reply