fable-seat

When the session model is Fable (claude-fable-5), apply this operating profile. On Opus — or any other non-Fable model — this entire profile is a NO-OP: apply none of it.

Fable is priced above Opus, and its mechanics reward different defaults than the models before it. This profile is the set of dials that make a Fable-led session run cheaply — nothing more. It is not a superiority claim. It does not assert Fable beats Opus; the standing no-measured-edge prior holds. The only claim here is narrow and conditional: if you're in the Fable seat, here's how to run it without paying for reflexes tuned to older models.

Gate first — is this session Fable?

Check the session model (stated in your environment / system context; the exact id is claude-fable-5).

• Not Fable (Opus, Sonnet, Haiku, anything else) → STOP here. This profile does not apply. Do not step effort down, do not de-prescribe your task framing, do not change caching behavior on account of this profile. It is a genuine no-op — the session runs exactly as it would if this file didn't exist.
• Fable → apply the dials below.

This gate is the whole safety of the profile. Every dial that follows is Fable-mechanics-specific, so applying one to a different model is a regression, not a neutral act.

The dials

1. Effort: default high, step DOWN — not xhigh by reflex

Fable inverts the old instinct. Its low/medium effort often clears what prior models needed xhigh to reach, and pushing effort up on routine work makes it over-gather and over-deliberate — you pay more tokens for worse-shaped output. So:

• Default to high for genuinely hard reasoning.
• Step DOWN to medium / low for routine judgment, mechanical edits, and well-scoped lookups. Most work lives here.
• Reserve xhigh / max for the hardest correctness-critical calls only — a subtle proof, a security-sensitive decision, a gnarly root-cause. Reaching for xhigh by default is the exact anti-pattern this dial exists to break.

2. De-prescribe the seat's OWN task framing — lead with the outcome

Fable does better with goal + constraints + the outcome you want, and worse when handed step-by-step scaffolding written for prior models — spelling out the steps drops its output quality. So when you frame a task — for yourself, or in the prompt you write for a subagent / child session:

• State the goal and the hard constraints, then say "lead with the outcome." Let Fable choose the path.
• Don't pre-chew the steps. The detailed how-to scaffolding that helped older models is a tax here.

Scope guard — this de-prescription is ONLY for the seat's own task framing. It does NOT license stripping any skill's safety rules, verification gates, confirmation gates, or the hard walls. Those stay verbatim. You are loosening the how-to scaffolding you author, never a guardrail a skill enforces.

3. Cache discipline — farm out, never inline-switch the main model

Prompt caching is a prefix match: the cached prefix is reused only while the system prompt and the leading context stay byte-identical. Switching the main loop's model mid-session — an inline model= swap — and back invalidates the whole cached prefix, so you re-bill the context on both switches. So:

• Keep the main loop's system prompt frozen and the tool set stable. Don't churn either mid-session.
• Route cheaper work to a subagent / child session, not to an inline model swap on the main loop. A child runs its own (cheaper) model without touching the main loop's cached prefix.

4. Task budget on long agentic runs

On a long, many-step agentic run, set a task / token budget so the run has a ceiling and can't quietly balloon. This is the cost guard for runs whose length you can't fully predict up front — pair it with dial 5 on the sessions that also run long.

5. Context editing / compaction on long sessions

Fable's context window is large, which means that left alone, a long session fills it with stale tool output and superseded turns that get re-billed on every step. On long sessions, use context editing / compaction to drop spent content, so you're not paying to re-read history that no longer matters.

6. Point at the durable memory surface

Fable re-derives less when it can read state instead of reconstructing it. Point it at the durable surfaces:

• the memory / notes file for facts that outlive the turn, and
• a per-task why — the goal + constraints — it can re-read instead of re-inferring.

Writing the durable thing down once beats paying Fable to reconstruct it every step.

What this profile never touches

• Safety rules, verification gates, confirmation gates, and the hard walls stay verbatim — on Fable and on every other model. Dial 2 loosens only the task-framing scaffolding you write, never a guardrail a skill enforces.
• No model comparison, no superiority claim. If you catch yourself writing "Fable is better / worse than Opus at X," cut it — that's outside this profile's scope and against the standing no-edge prior.

One-line no-op restatement

Not on Fable? None of the above applies — run the session exactly as you otherwise would.