Published: 2026-06-23
Analysis & perspective

Task Imagination: The Skill Big Models Like Fable 5 Demand

Chapters / key moments (click to jump — plays here on the page)

Nate B Jones argues the real constraint with frontier models like Fable 5 isn't capability — it's our ability to imagine a big enough ask. He makes the case for "task imagination": handing a model whole, ambiguous jobs (not tracker-sized tasks), writing down what "done" looks like, assembling a data pack, then walking away and reviewing the result like an owner. This is analysis and perspective, not a step-by-step guide.

Source video

"Task Imagination is the New Skill. Here's Why Claude Fails" by Nate B Jones — Watch on YouTube →

Key Takeaways

The bottleneck flipped. With a big model, the limit Jones kept hitting wasn't the model running out of ability — it was running out of big things to ask. If your asks stay prompt-sized, every frontier model feels roughly the same.
"Task imagination" ≠ delegation. Delegation is tasks already on your tracker with a name attached. Task imagination is the dirty, ambiguous jobs nobody has written down — de-duping 2M CRM records, fact-checking a 500-page board packet, mining 40,000 reviews — because they felt too big to assign.
Define "done" first. Write a clear paragraph describing what should exist at the end, assemble a data pack (this can take hours), hand it over — and then do the hard part: walk away and stop hovering. The itch to babysit is a habit trained on models that used to be too small.
Review like an owner. When the work comes back, check it like a senior stakeholder's output: is the scope right, is it accurate, is it angled correctly — then assign revision work as needed.
Reserve big models for big jobs. At ~$50 per million output tokens, Fable 5 isn't a daily driver. Spending that on a summary a cheap model could do in seconds wastes the muscle — use it for serious work where one job can save weeks.
On jobs: the roles most exposed are pure, zero-judgment execution. For everyone else, the shift is toward being a "model manager" — scoping, feeding data, and judging output — which Jones frames as a career-uplift opportunity, not a layoff sentence.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Find the job nobody has written down
List the gnarly, ambiguous jobs on your desk that are too big to have ever been assigned: fact-checking a 500-page board packet, merging and de-duplicating millions of CRM records, or reading 40,000 reviews. He says if it makes you sigh, it qualifies.
Write what done means, in one paragraph
Describe the finished thing you want to exist. Then build a data pack with all the source material the model needs; he says to budget a few hours for this.
Give it, then walk away
Hand over the pack plus the done-paragraph and rough guidelines for judgment calls. Don't hover.
Review like an owner
Check the result is correct, angled right and covers the full scope, and assign revision work for gaps.

Gotchas

He calls it too expensive to be a daily driver and says it is overkill for email-sized tasks ($50 per million output tokens by his figure in June 2026; check current pricing).
What he saw go wrong: visual design wasn't one-shot quality (clipped headings, charts a designer would improve), it missed information that existed only in handwritten images until told to look, and every run still ended with review work for him.
What he praised: it quarantined bad data rather than silently fixing it, inventoried fake credentials without leaking them, and built a review queue of the calls it was unsure about.