For most of the past three years, the path to getting better at AI was learning to ask smaller. Break the task into steps. Be specific. Do not overwhelm the model. That discipline served practitioners well — until the same habits started producing worse results on the newer, more capable models, not better ones.
Nate’s review of Claude Fable 5 named something many practitioners had been sensing but could not quite articulate: we have an ask-size problem, and it works in both directions. We trained ourselves to ask small because bigger asks failed in 2023–2024. That was a rational adaptation at the time. It is now a liability.
Fable 5 — Anthropic’s frontier model, briefly available before being pulled from production following a US government review — was built for hours-long work. Not summaries, not rewrites, not quick code snippets. Jobs. Whole, messy, multi-file jobs that previously took a person a day or a week. If that is true, then prompts written for smaller models are actively underperforming on this one. The five prompts in Nate’s piece are designed to help close that gap — specifically, to rebuild the thinking habits you will need when the model comes back.
The first thing to fix: your existing prompts
The first prompt in the set is a Prompt Auditor. You paste in an existing prompt or saved instruction set, and it diagnoses every small-model habit it finds: chunking instructions (“do this in parts”), context re-paste rituals, role-play incantations (“you are an expert analyst”), over-constraint, single-question framing where a whole job is actually implied. Then it rebuilds the prompt for what the article calls whole-job handoff.
This proves useful as a diagnostic lens even before running it formally. Scanning saved prompts surfaces exactly the patterns described: narrow step-by-step instructions that assumed the model could not be trusted to manage its own work, and artificial scope ceilings that made sense for earlier tools and are now just friction.
The rebuilt prompt targets a nine-field briefing structure — a “Whole-Job Spec” — that names the finished artifact, points at the source material, sets boundaries (what to touch, what to leave, what needs a human ruling), states a review standard, and asks the model to leave a proof trail: a working log of its judgment calls and uncertainties, left for human inspection. The framing is “the briefing you would give a capable person on their first day.” That has stuck with me.
The second thing to fix: context that reads as risky
The second prompt is a Guardrail Pre-Flight checker. Fable 5 carries additional safety measures around dual-use capabilities — dual-use meaning content or techniques that have both legitimate and harmful applications, such as security research or biology — and legitimate professional work can appear risky on a careful read if the benign context is not stated explicitly.
The prompt helps you make your honest professional context legible before the run: your role, your authorization, the defensive or educational purpose of the work. This matters more than it might seem. The way a security professional describes their work to a colleague is different from how that same work reads to a model trained to be cautious. The gap is not about intent — it is about legibility.
The shift underneath all of it
Both prompts are in service of something the article calls detailed task imagination — the ability to see ambient, unowned work (the review that never gets written, the CRM list that never gets cleaned, the migration that stays on the backlog) as a job with a finish line and hand it over with full context.
Nate frames this as the learnable skill that actually unlocks a model like Fable 5. Not better prompt syntax. Not smarter role definitions. The ability to imagine, concretely, what “done” looks like for a piece of work that currently exists only as background noise.
Seven questions help build that imagination: What finished artifact should exist? What messy source material is needed? What tools would a human open to do this? What judgment calls arise along the way? What should be left untouched? What would convince you the job is done? What evidence should be left behind? Those make a useful quick checklist before any significant run — not just Fable-sized ones.
Why prepare now, while it is not available
Fable 5 was pulled. The review exists because the capability direction it represents is durable, not because the model is currently in production. That seems like exactly the right reason to do the preparation now, while there is no cost pressure and no impatient deadline.
We tend to learn new tools reactively — when the pressure is on and the run is already in progress. Auditing your existing prompts, running the pre-flight on anything that touches sensitive domains, and practicing whole-job thinking on small, cheap runs first: that is the kind of preparation that compounds, whatever form the next capable model takes.
Text summarized and optimized using Anthropic’s models and reviewed by a human.