Prompting fable 5

Prompting Claude Fable 5 (and Mythos 5)

Fable 5 is the top Mythos-class tier (above Opus). It shines on long-horizon, ambiguous, hours-to-days work. Prompt/scaffold it differently from Opus 4.8. Source: Anthropic's Prompting Claude Fable 5 guide

• hygiene lessons from benign review work that kept falling back to Opus (snapshot 2026-07-02).

⚠️ Why benign work sometimes falls back to Opus (read this FIRST)

Fable 5 runs safety classifiers. When one fires, the turn returns stop_reason: "refusal" and falls back to Opus 4.8 — in some Claude Code configurations that swap happens automatically. That fallback is the intended safety net, and most of the time it is the system working correctly. This section is only about one narrower thing: the classifiers also produce false positives on legitimately benign work, and when honest, in-scope work keeps bouncing it's worth knowing why and how to cut the noise.

Scope — this is about false positives on benign work, nothing else. If a task is genuinely in a restricted domain (real exploit/malware development, actual dangerous-bio methods), the fallback is the intended safety behaviour — that's the classifier doing its job, and you should not try to circumvent it. Everything below applies only when ordinary, legitimate work trips the filter. If you're unsure which case you're in, treat it as the intended behaviour and let the fallback stand.

Where benign work can trip it:

1. Security review of your own code — a read-only audit (RLS, auth, cross-tenant isolation, secret handling, IDOR) can score like offensive tooling even though you're reviewing an app you own.
2. Biology / life sciences — benign lab or mechanism discussion can trip the bio classifier.
3. reasoning_extraction — any instruction that tells the model to echo / transcribe / explain its internal reasoning as response text. This is the sneakiest, because it's self-inflicted and has nothing to do with your topic — a "show your thinking" instruction trips it on otherwise-benign turns.

The two self-inflicted triggers — both are simple hygiene fixes

• "Show your thinking" output styles or instructions. The Claude Code "learning" and "explanatory" output styles inject "★ Insight" blocks / "explain your implementation choices" on every turn → they trip reasoning_extraction on EVERY Fable turn, even benign ones. Turn these output styles OFF for Fable. Audit skills/system prompts for "walk me through your reasoning", "explain your thought process", "reflect and show your work" and remove them — don't instruct the model to echo its raw reasoning as prose. If you need reasoning visibility, read the structured thinking blocks (adaptive thinking) instead, or use a send-to-user tool for progress. (Quick check: if your first reply contains a "★ Insight" block, the output style is still on.)
• Keep durable state in files so a fallback loses no work. A fallback mid-run shouldn't cost you anything. Persist work to files and run long jobs as background workflows/tasks, so if a turn falls back to Opus the state is already saved and you just continue. This is good hygiene regardless of classifiers.

If benign review work keeps falling back

• Describe the task as what it honestly is — "review my own app's tenant isolation", "readiness assessment" — because that is the accurate description, not to slip anything past a check. Accurate scoping won't (and shouldn't) get restricted content through; it just avoids mislabelling benign work as offensive.
• Prefer fresh-context verifier subagents for review — they beat self-critique on quality anyway (see below). But a subagent's model is configuration-dependent: a subagent that inherits Fable will hit the same classifier and fall back too. Verify the subagent config rather than assuming subagents are immune.
• If genuinely-benign work still keeps bouncing, treat it as a known false-positive limitation — accept the Opus fallback for that work, or report the false positive through normal feedback channels. Don't re-architect the task specifically to keep flagged content away from the classifier; that would cross from "reduce noise on benign work" into working around a safety check, which is out of scope here.

Effort is the primary control

• Default high for most tasks. xhigh for the most capability-sensitive work. medium/low for routine/mechanical work — Fable's low still often beats prior models' xhigh.
• Lower effort if a task finishes but takes longer than needed, or you want a snappier interactive loop.
• Effort tracks reversibility/blast-radius, not task size. In Claude Code, the user owns the effort knob (/model, /fast) — flag or delegate; don't silently run an expensive phase at low effort or vice-versa.

Longer turns by default — restructure for async

Hard requests run many minutes; autonomous runs, hours. Adjust timeouts; check on runs asynchronously (background workflows/tasks that notify on completion) rather than blocking. Don't poll harness-tracked work.

Strong instruction-following — steer with brief instructions, not enumerations

Fable follows short directives well; you don't have to name every behavior. Useful stock instructions:

• Brevity/readability: "Lead with the outcome. Being readable and being concise are different things; readability matters more. Don't compress into fragments, arrow-chains, or jargon."
• Checkpoints: "Pause for the user only when the work genuinely requires them — a destructive/ irreversible action, a real scope change, or input only they can provide. Then ask and end the turn."
• Ground progress claims: "Before reporting progress, audit each claim against a tool result from this session. Only report work you can point to evidence for; if unverified, say so." (Near-eliminates fabricated status.)
• Scope discipline: "Don't add features, refactors, abstractions, or defensive error-handling beyond what the task requires. Do the simplest thing that works."
• Anti-overplanning (Fable over-deliberates on ambiguity at high effort): "When you have enough information to act, act. Don't re-derive facts already established, re-litigate a settled decision, or narrate options you won't pursue in user-facing messages. If weighing a choice, give a recommendation, not an exhaustive survey. (This doesn't apply to thinking blocks.)"
• Boundaries: "When the user is describing/asking/thinking-out-loud rather than requesting a change, the deliverable is your assessment — report and stop; don't apply a fix until asked."

Parallel subagents + memory + send-to-user

• Delegate readily. Fable dispatches and sustains parallel subagents well. "Delegate independent subtasks and keep working while they run; intervene if one goes off track." Long-lived subagents keep context (cache reads) and avoid bottlenecking on the slowest. Prefer async over blocking.
• Self-verification: fresh-context verifier subagents beat self-critique. "Verify your work with subagents against the spec at intervals." Note a subagent's model is configuration-dependent — it can run on a different configured model, but a subagent that inherits Fable behaves like Fable (same effort characteristics, same classifiers); check the config rather than assuming otherwise.
• Memory: give it a place to write lessons (one file per lesson, one-line summary on top; update, don't duplicate; delete wrong notes). Bootstrap by having it reflect on past sessions via subagents.
• send-to-user tool (long async agents): a client tool that renders its input verbatim without ending the turn — for deliverables/direct answers the user must see exactly. Pair with an elicitation instruction or it won't get called. Don't route narration through it.

Rare late-session quirks (and the one-line fixes)

• Early stop (text-only "I'll now run X" with no tool call): "continue" / "go ahead end-to-end" suffices; add an autonomous-operation reminder for pipelines.
• Context-budget worry (offers to hand off / summarize): avoid surfacing token countdowns; if you must, add "You have ample context remaining; do not stop or suggest a new session on account of context limits."

Scaffolding changes when migrating from Opus

• Start at the top of your difficulty range — give it your hardest unsolved problem, not simple workloads.
• Refactor over-prescriptive skills/prompts — Opus-era enumerated instructions can degrade Fable's output. Remove older micro-instructions when default behavior is already better. Fable updates skills on the fly.
• Never instruct it to reproduce/echo its reasoning in the response — triggers reasoning_extraction fallbacks (see the gotcha section). Audit for reflection/show-your-thinking language when migrating.
• Give the reason, not just the request — "I'm working on X for Y; they need Z; with that in mind: …".

One-line diagnosis for "my BENIGN Fable session keeps dropping to Opus"

If genuinely benign work is falling back, check in order: (1) an active learning/explanatory output style or any show-your-thinking instruction → turn it off (the highest-yield fix, and pure hygiene); (2) your prompts/skills contain "echo/explain your reasoning" language → remove it. If neither applies and the work is truly benign, this is a known false positive — accept the Opus fallback or report it, and make sure your state is in files so the fallback costs you nothing. Reminder: if the work is actually in a restricted domain, the fallback is the intended behaviour, not a bug to route around.