just-works: Skill for Claude Code

Fable 5 Prompting

When to Use

• Creating or editing system prompts targeting Claude Fable 5
• Designing long-running autonomous agents, harnesses, or subagent orchestration on Fable 5
• Migrating prompt text from Opus 4.8 or older Claude models
• Diagnosing refusals, fallbacks, early stopping, or fabricated progress reports on Fable 5

Overview

Claude Fable 5 is Anthropic's most powerful model — a new tier above Opus (1M context, 128k max output, $10/$50 per MTok). It takes on problems previously too complex, long-running, or ambiguous for prior models, and is particularly effective at end-to-end work that takes a person hours, days, or weeks. Teams seeing the best outcomes apply it to their hardest unsolved problems; testing it only on simpler workloads undersells its capability range. It also performs reliably on straightforward tasks.

API surface matches Opus 4.7/4.8 (adaptive thinking only, no sampling parameters, no budget_tokens, no last-turn prefills) with one new breaking change: an explicit thinking: {type: "disabled"} returns a 400 — omit the thinking parameter entirely instead. The upstream guide covers Claude Fable 5 and Claude Mythos 5; this skill targets Fable 5 (claude-fable-5).

Capability improvements at this level are a prompt to re-evaluate which instructions, tools, and guardrails are still needed — skills and prompts developed for prior models are often too prescriptive for Fable 5 and can degrade output quality.

• Long-horizon autonomy: Sustains productive output over extended periods — multi-day, goal-directed runs with strong instruction retention.
• Longer turns by default: Individual requests on hard tasks can run many minutes at higher effort; autonomous runs extend for hours. The largest shift teams encounter when migrating.
• First-shot correctness: Single-pass implementations of well-specified complex systems that previously took days of iteration.
• Strong instruction following: A brief instruction steers most behaviors — no need to enumerate each pattern by name.
• Readier subagent dispatch: Dispatches parallel subagents more readily than prior models; dependable at sustaining communication with long-running subagents and peer agents.
• Stronger vision: Interprets dense technical images, web apps, and screenshots with substantially higher accuracy, often using fewer output tokens; trained to use bash and crop tools on flipped, blurry, or noisy images.
• Better code review and debugging: Noticeably higher bug-finding recall than Opus 4.8 (outside safety-classifier domains), including search across codebases and repo history.
• Navigates ambiguity: Performs well on complex, multi-threaded requests when asked to determine next steps.
• Memory leverage: Performs particularly well when it can record lessons from previous runs and reference them.
• Occasional unrequested actions: Can draft an email no one asked for or create defensive git-branch backups — state boundaries explicitly.
• Rare early stopping: Deep into long sessions, can end a turn with a statement of intent without issuing the tool call, or ask permission it doesn't need.
• Rare context-budget anxiety: In very long sessions can suggest a new session or trim its own work — usually triggered by a visible remaining-token countdown.
• Safety classifiers: Targets offensive cybersecurity, biology/life sciences, and extraction of summarized thinking. Benign work in those areas may also trigger refusals.