Claude Fable 5 Prompt Leak Is a User Manual for Long-Running Agents
The leaked prompt points to files, tools, artifacts, checks, and work packets instead of longer one-shot prompts
In ~6 mins: the system prompt leak, the jailbreak posts, the Fable 5 Workload Test, a Fable-ready task brief, and the tradeoffs developers should check first.
Claude Fable 5’s leaked system prompt is 1,585 lines, but the useful lesson is much shorter: Fable is built for long-running agent work, not routine short tasks.
The jailbreak posts made the leak spread fast, but the more useful question is why the prompt is so tool-heavy in the first place.
The answer is that Fable works best when it can inspect files, use tools, produce artifacts, and verify progress across a durable job.
What and Why
Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Within about a day, Pliny the Liberator posted what he described as the full Fable 5 system prompt to X and GitHub.
The next day, he posted a jailbreak thread with screenshots showing Fable 5 complying on high-risk tasks. Anthropic’s own system card frames this category more narrowly: its public bug bounty had about 100,000 attempts by June 5, with no universal jailbreak and two task-specific jailbreaks.
The practical story is not that everyone should copy a leaked system prompt. The practical story is that the prompt exposes the shape of the job Anthropic expects this model to perform.
It is full of rules for tools, files, search, artifacts, memory, citations, presentation, and safety. That reads less like a magic incantation and more like an operating manual for long-running agent work.
Context
A Reminder: Fable 5 is Anthropic’s generally available Mythos-class model with added safeguards. The important specs for this piece are simple: 1M context, 128k max output, and API pricing of $10 per million input tokens and $50 per million output tokens.
Anthropic’s own launch example gives the better frame. Stripe used the model to migrate a 50M-line Ruby codebase in one day, work Anthropic says was estimated at more than two months for an engineering team.
That is the kind of job where Fable makes sense. The model has enough room to inspect a lot of context, but the payoff comes from the work structure around it.
AlphaSignal’s earlier hands-on test reached the same pattern from a different direction: Fable looked best when difficult ML work had tests, labels, or held-out checks that could reject a convenient answer.
This piece asks the step before prompting: when does a task deserve Fable at all?
Core Concept
Use the Fable 5 Workload Test before spending Fable tokens.
Give the task to Fable only if it has at least three of these traits:
Multiple stages.
Repository, file, document, or dataset inspection.
Tests, evals, citations, screenshots, or another way to check the work.
A durable output: patch, report, artifact, memo, notebook, PR notes, or migration plan.
Long-context reasoning across many sources.
Visual, document, or browser review.
Async checkpoints where the model can report state and keep going.
A high cost for being wrong.
If the work can be finished correctly in one short turn, do not use Fable. If the work needs context gathering, execution, verification, and a reusable output, Fable becomes easier to justify.
The bad use case is not “easy work.” It is work where the model has no room to improve the answer by inspecting, acting, and checking.
That means Fable is a poor default for renaming a variable, rewriting a paragraph, summarizing a short doc, or producing five quick ideas. Those tasks do not need a long-running agent.
Fable-shaped work looks different. It asks the model to analyze a repo-wide bug, patch it, write the regression test, run the checks, and explain the remaining risk.
It also fits research work with a finish line. Give it papers, code, notes, and an output format, then ask it to produce a decision table with citations and gaps.
Before using Fable, ask five questions:
Can the team define what done means?
Can the model inspect the relevant files or sources?
Can the model run or receive verification?
Will the output be reused after the conversation?
Would a wrong answer cost more than the model run?
Three yeses is the entry point. Fewer than three usually means a cheaper model should take the first pass.
Technical architecture or mechanics
The leaked prompt reads less like one perfect instruction and more like a workbench inventory.
Across the prompt, the repeating surfaces are tools, files, web search, artifacts, memory, citations, refusal boundaries, and answer formatting. The lesson is not “write a longer prompt.” The lesson is “give the model a better work packet.”
A Fable-ready task brief should include:
Goal: what should exist at the end.
Context and files: repo paths, docs, tickets, datasets, links, or prior decisions.
Constraints: what the model must not change, assume, or expose.
Acceptance criteria: how success will be judged.
Verification steps: tests, commands, reviewers, evals, screenshots, or citations.
Deliverables: patch, report, artifact, table, PR notes, or decision memo.
Checkpoints: when to pause, summarize, or report unresolved risks.
That brief is not a giant prompt. It is a control surface for long-running work.
For example, “fix this flaky test” is too small for Fable unless the failure is deep. “Trace this flaky integration test across the repo, isolate the nondeterministic dependency, patch the code, add a regression test, and report the exact commands run” is closer to a Fable job.
The prompt leak also explains why old prompt habits can backfire. If the model already has strong defaults for planning, tool use, and verification, a wall of step-by-step instructions can turn into friction.
For prompt-level cleanup, AlphaSignal’s earlier Fable prompting guide covers the official Anthropic advice on deleting old instructions, setting effort, and adding verification blocks. This article sits one layer above that: deciding which jobs should reach Fable in the first place.
One caveat: the leaked prompt appears to describe the chat-app surface. API use, Claude Code, Cursor, and other agent harnesses may expose different tools and policies.
Evidence
The official positioning supports the workload test. Anthropic describes Fable as a model for long-running, complex, asynchronous work, and the Stripe migration example is the cleanest proof point in the launch material.
The leaked source supports it too. A 1,585-line prompt spends less space teaching facts and more space defining the operating environment around the model.
That is the part developers can use. The takeaway is not to copy leaked instructions, but to notice the surfaces Anthropic gives the model: tools, files, memory, artifacts, search, citations, and safety rules.
AlphaSignal’s previous ML testing adds the verification lesson. Fable performed best when the task had a measurable finish line, and the weaker moments appeared when the task could satisfy public checks while still failing a private objective.
The jailbreak evidence should be read with care. The screenshots show task-specific bypasses, while Anthropic’s system card says its broader bug bounty found no universal jailbreak before launch.
The enterprise evidence matters because the strongest model is not always the easiest model to adopt. Anthropic says Mythos-class prompts and outputs are retained for 30 days for trust and safety, and The Verge reported that Microsoft limited employee access while legal teams reviewed the policy.
Safety routing is another adoption detail. Anthropic says safeguard-triggered fallback affects fewer than 5 percent of sessions, but the company also describes the classifiers as conservative, with false positives expected.
SOTA Context
The useful comparison is simple: cheaper models for routine turns, Fable for durable jobs, and human review or fallback plans for sensitive domains. If the task does not need context gathering, execution, verification, and an artifact, Fable is probably the wrong first model.
AlphaSignal Take
Here’s what the leak does not settle: whether Fable is worth using depends less on raw capability than on task packaging.
Our read is that Fable’s edge shows up when the work packet has files, constraints, tests, and a reusable output. Without those, teams are paying premium rates for a model that has no real room to operate.
The counterargument is real. Cost, 30-day retention, fallback behavior, and false positives can make Fable wrong for teams with strict data rules, sensitive research, or high-volume routine workflows.
So the best recommendation is to reserve Fable for work where verification and artifact value justify the spend. The right question is whether the job is durable enough, not whether the prompt is long enough.
Who benefits and who doesn’t
Fable fits staff engineers running migrations, AI engineers building long-running agents, ML engineers debugging production model failures, and research teams turning papers, code, and data into checked artifacts.
It is the wrong default for short coding chores, high-volume summaries, generic brainstorming, routine copy edits, and organizations that cannot accept the retention or fallback tradeoffs.
Practitioner Implication
Package long-running work packets for engineering and research teams now that Fable can hold context, use tools, and verify outputs across multi-stage agent runs.
Links
Anthropic launch post (release and safeguard details, ~8 min read)
Claude models overview (model specs and availability, ~4 min read)
Claude Fable 5 and Mythos 5 system card (safety and eval details, PDF)
Mythos-class data retention practices (30-day retention policy, ~4 min read)
Fable 5 prompt leak post (original X post)
Fable 5 jailbreak post (original X post)
CL4R1T4S Fable prompt archive (prompt archive)
Claude Fable 5’s Hardest Test Was Knowing the Metric Was Wrong (AlphaSignal hands-on testing)
How to Actually Prompt Claude Fable 5 (AlphaSignal prompting guide)
The Verge on Microsoft access limits (enterprise reaction)
WIRED on Anthropic safeguard backlash (safeguard controversy)
Follow @AlphaSignalAI for more content like this.
Subscribe at alphasignal.ai/newsletter for daily AI signals. Read by 300,000+ subscribers.
Questions?
Q: What did the Claude Fable 5 system prompt leak reveal?
A: The leak showed a long prompt organized around tools, files, search, artifacts, memory, citations, formatting, and safety rules. The useful lesson is that Fable expects a full operating environment rather than a longer user prompt.
Q: What is Claude Fable 5 best used for?
A: Fable 5 is best used for long-running engineering and research tasks with files to inspect, actions to take, and verification to run. It fits migrations, multi-file debugging, research synthesis, and artifact-heavy technical work.
Q: Should developers use Claude Fable 5 for small coding tasks?
A: Usually no. If a cheaper model can finish the task correctly in one short turn, Fable’s long-context and agentic strengths are mostly unused.
Q: How should teams prompt Claude Fable 5 after the leak?
A: Teams should package the task as a work packet with a goal, files, constraints, acceptance criteria, verification steps, deliverables, and checkpoints. The leaked prompt is not a template to copy, but it shows which surfaces matter.
Q: What are the main risks of using Claude Fable 5?
A: The main risks are cost, 30-day Mythos-class retention, fallback behavior, and false positives near sensitive domains. Teams working near cyber, bio/chem, distillation, or frontier-model topics need a clear fallback and review plan.