Anthropic's Fable 5 Prompt Library: A Framework for Delegating Long-Running Work to Claude
July 6, 2026 · AI Automators
If you build automations, that framing is the interesting part. Most of us reach for models that answer in seconds. Fable 5, as described here, is pitched as the opposite — a model you set up carefully, walk away from, and check on later.
What the library actually is
The pack centers on a simple sequence: plan, build, research, verify, and hand off complex work. According to the page, Fable 5 is positioned by Anthropic as a strong coding model for ambitious projects, and it's most useful inside Claude Code, where it can inspect your files, call tools, orchestrate subagents, and check its own output.
The prompts map to concrete jobs rather than abstractions. Examples include: run a task overnight, plan a product or architecture, verify that a build works, port a codebase, fix a broken agent workflow, turn feedback into a batch of changes, build a first version from a spec, and organize agent context. There's also an anchor interview — Dan Shipper talking to Mike Krieger, Instagram cofounder and head of Anthropic Labs — that four of the templates are drawn from. Both the full prompt pack and the transcript are downloadable as Markdown files.
Two prompts stand out because they're about restraint rather than doing more work. The first, "Find Fable-worthy work," explicitly tells the model *not* to execute anything. Instead it inspects your real context — repos, docs, Slack, Linear, analytics — and scores candidate tasks from 1 to 5 on things like multi-source context, delegation fit, judgment required, clear finish line, and leverage. It then down-ranks anything short, obvious, highly interactive, or better handled by a human, and packages the best three as ready-to-run "Fable Briefs."
The second, "The Fable Brief," is the assignment format itself: the problem to solve, the final outcome, sources to inspect, constraints, suggested workflow, human checkpoints, and the evidence required before the task is considered done.
When it's worth the wait
The library is unusually honest about cost. It spells out that a long Fable run is slower and more expensive, and gives a decision rule. Use Fable when the job pulls from several sources or tools, can keep moving without constant input, and has a finish line you can describe and test. Use Codex or a faster model when you expect to steer every few minutes, the task is short with an obvious path, or a long run would cost more than the result is worth.
That's the right instinct for anyone building automations. The failure mode with capable agents isn't that they can't do the work — it's that people point them at tasks that need constant supervision, then spend more time babysitting than they saved. The "describe and test the finish line" criterion is the load-bearing one. If you can't state what a correct result looks like and how you'll check it, you can't safely delegate it to anything that runs for hours.
There's also a setup component. The pack points to a "Compound Engineering" plugin that Every uses daily, which is meant to guide a full brainstorm, plan, build, review, test, and pull-request pipeline, and to give the model a structured way to record lessons after a run. That last part — saving what the agent learned — is one of the 13 jobs, and it's the kind of thing most people skip and later regret.
A fair caveat: much of the value here rests on claims from Anthropic and Every about how well Fable 5 performs on ambitious coding tasks. The page calls it "the best coding model in the world," which is a marketing statement, not a benchmark you should take at face value. The prompts, however, stand on their own regardless of which model you run them against. The Fable Brief structure and the pre-work triage would improve delegation to almost any capable agent.
Where it fits for builders
If your automation stack today is trigger-and-response — Zapier, Make, or n8n firing short model calls via OpenAI or Claude — this library covers a different layer. Those tools are for fast, deterministic steps. Fable-style work is for the messy, judgment-heavy jobs that don't fit a linear flow: porting a codebase, debugging a broken agent, or turning a pile of feedback into a batch of concrete changes.
The practical move is to treat the triage prompt as a filter. Run it against your actual backlog, see what scores well, and only escalate the handful of tasks that genuinely justify a long, autonomous run. Everything else stays with a faster model or a person.
The pack is free to download, so the cost of evaluating it is low. If you want help wiring these delegation patterns into your own build-and-verify pipeline, browse the provider directory to find someone who can put it to work.