Claude fable skills

Claude Fable

Adapt prompts and scaffolding to Claude Fable 5's behavior rather than reusing older Opus-era instructions unchanged. Keep the core skill lean and load the detailed reference when you need examples or exact wording.

Workflow

1. Identify whether the task is about prompt design, harness design, agent behavior, or migrating older instructions to Claude Fable 5.
2. Read references/prompting-patterns.md when the task needs exact behavioral guidance, sample instruction blocks, or migration detail.
3. Rewrite the user's prompt, skill, or system guidance to match Fable's strengths: high autonomy, strong instruction following, longer turns, better repo search, and better subagent delegation.
4. Remove over-prescriptive or outdated instructions that were compensating for weaker earlier models.
5. Keep the final output operational: give the revised prompt, scaffold, or skill text directly, and note only the key rationale that affects behavior.

Core Patterns

Scale the task up

Start with harder, more end-to-end tasks than you would have assigned to older models. Fable performs best when given meaningful autonomy over complex work instead of only narrow toy prompts.

Tune effort deliberately

Use effort as the main latency vs. quality control. Default to high, reserve xhigh for capability-sensitive work, and step down to medium or low when the task is routine or overly slow.

Bias toward action once context is sufficient

Fable can overplan on ambiguous tasks. Add short instructions that tell it to act once it has enough information and to avoid surveying options it will not pursue.

Use short steering, not long behavior checklists

Fable usually responds well to compact instructions for brevity, checkpointing, and user communication. Prefer a small number of strong directives over enumerating every anti-pattern.

Constrain unrequested work

For focused implementation tasks, explicitly forbid bonus refactors, speculative abstractions, unnecessary validation, and hypothetical future-proofing.

Ground progress in evidence

For long runs, require every reported status claim to map to a tool result from the current session. This reduces fabricated progress updates.

Define pause boundaries

Tell Fable to stop only for destructive actions, real scope changes, or missing user-only input. Otherwise it should keep going.

Encourage subagent parallelism

Fable is good at delegating independent subtasks. When building orchestration prompts, tell it to dispatch independent work to subagents and continue working while they run.

Give it memory hooks

If the environment supports persistent notes, define a simple memory system and instruct Fable to store one lesson per file with a one-line summary.

Improve asynchronous UX

If the agent runs for a long time, pair it with a send_to_user tool or equivalent client-side message tool. Instruct the model explicitly when to use it for verbatim user-facing content.

Migration Guidance

• Replace legacy prompts that force the model to expose or restate its internal reasoning.
• Revisit older skills that are too prescriptive; simplify them unless the task is genuinely fragile.
• Add explicit autonomy instructions for unattended runs so the model does not end on promises or unnecessary permission checks.
• Avoid showing explicit context-budget countdowns when possible; if unavoidable, reassure the model that it has enough context remaining.

Reference Map

Read references/prompting-patterns.md when you need:

• effort-selection guidance
• sample instruction blocks for brevity, checkpointing, autonomy, or progress reporting
• migration notes from Claude Opus 4.8 to Claude Fable 5
• memory-system, subagent, or send-to-user tool guidance
• caveats around refusals, cybersecurity, biology, or reasoning-extraction behavior