fable-safe-authoring

Fable-safe authoring

Skills written for older models degrade frontier output. Treat the rules below as hard constraints when authoring anything in this marketplace — the goal is output that holds up on Fable 5 without tripping a refusal or burning tokens on over-prescription.

Write contracts, not recipes

Each skill, command, or agent is: the goal, the rationale behind it, the explicit boundaries (what is out of scope), the self-verification criteria (how the model knows it succeeded), and an explicit marker for where to pause for human input. Do not transcribe a step-by-step procedure of what the model should do internally — over-prescription measurably degrades frontier output and wastes tokens. State the outcome and the constraints; let the model find the path.

Ask for evidence, never introspection

Never instruct a skill, command, agent, or dispatch prompt to narrate, transcribe, or justify its internal thinking — phrases in the shape of explain your reasoning, show your work, or think out loud. On Fable 5 these trip the reasoning_extraction refusal category: the response comes back with stop_reason of refusal, and Claude Code silently falls back to Opus 4.8 — so the kit quietly stops running on the model you targeted.

Ask instead for evidence: the diff, the command output, the failing-then-passing test, the file:line. Evidence is verifiable and model-safe; introspection is neither.

Handle a refusal as a first-class outcome

When wiring any code that consumes model output (the release sidecar, a hook, a generated app's API client), branch on stop_reason == "refusal" and treat it as a non-result — surface it, retry with a reworded prompt, or fall back deliberately. Do not treat a refusal as valid content.

• Branch on stop_reason, not on stop_details. stop_details ({ type: "refusal", category, explanation }) is informational only: category and explanation can be null, and stop_details is absent entirely on the Message Batches API. The refusal categories include reasoning_extraction, cyber, bio, and frontier_llm.
• There is no API knob to disable or expand thinking on Fable 5. Its adaptive thinking is summarized-only and cannot be turned off; showThinkingSummaries is a Claude Code CLI setting, not a request parameter. Do not author any skill that assumes a thinking display/expansion request field exists.

Lead with the outcome, stay terse

A subagent's final message is data for the orchestrator, not prose for a human — structure it outcome-first (what changed, then evidence). A user-facing summary re-grounds the same way: the result first, the path second. Terseness is a feature; padding is tokens the orchestrator has to read.

Self-check before shipping a skill/agent/command

• No instruction to explain, narrate, or justify reasoning — anywhere.
• Every "do" is paired with a verifiable signal of done.
• Agent model: is an alias (sonnet / haiku / opus / fable / inherit), never a dated id.
• The description is the trigger ("what + when"), tight enough to route on.