Fable 5 works for hours at a stretch. Your job is now the checkpoint prompt.
The defining claim of Claude Fable 5 isn't a benchmark score. It's a shape: the longer and more complex the task, the larger its lead. Stripe's 50-million-line migration story, the multi-hour autonomous runs, the millions-of-tokens focus: they all point the same direction.
That direction changes what a human in the loop is for.
The old loop and the new one
With minute-scale completions, the loop was: read the answer, judge it, fix it, re-prompt. Your value was per-answer quality control, dozens of times a session.
With hour-scale autonomy, you intervene a handful of times, and each intervention steers a much larger block of work. Get a checkpoint wrong and the model confidently compounds the error for the next hour at $50 per million output tokens.
So the question we care about, the one this site exists for, gets more important, not less: at a checkpoint, what do you actually say?
What our data already says
We've run 184 blind judgments on intervention prompts. The mechanism findings transfer to checkpointing better than the individual prompts do:
Forcing functions beat scenarios. "Now tear that apart. Where is it weakest?" consistently won; "trace what happens if two users collide at the same millisecond" consistently lost. At a checkpoint on a long autonomous run, you don't know what the model did in detail, which is exactly when an open forcing function outperforms a specific guess. (The full story of losing to "are you sure?" is required reading here.)
Premature-complexity pushback has a niche. Our experimental subtract category ("What complexity here is paying rent? Evict the rest.") won precisely on the failure mode long autonomous runs are most prone to: accumulating structure nobody asked for. A model that can do two months of work in a day can also over-engineer at unprecedented speed.
Escalation matters. Pressing the same concern twice should get sharper, not louder. That's the escalation ladder in the deck: gentle on rung 1, surgical by rung 4.
What we haven't tested
Honesty section: every judgment above was collected on pre-Fable models. Whether category-matched checkpoints beat a generic "review your progress" on Fable 5 specifically is an empirical question, and our track record says intuition is a bad oracle; our first hand-curated deck lost to a three-word baseline. We're re-running the eval harness against claude-fable-5 and will publish the table either way.
Until then: the deck's core tier is the set that survived blind judging, and Space is still the fastest way to get an intervention that isn't "looks good, continue."
Keep Working is a side project of Urbano DX. The day job is building AI software for clients: paid proof-of-concepts that de-risk exactly this kind of "what does the new model actually change for us" question, plus production APIs and AI workflows. If you'd rather measure than guess, we should talk.